Agent package + coding agent WIP, refactored web-ui prompts

docs/agent.md

@@ -1,15 +1,23 @@
-# Coding Agent Architecture
+# Agent Architecture
 
 ## Executive Summary
 
-This document proposes extracting the agent infrastructure from `@mariozechner/pi-web-ui` and `@mariozechner/pi-agent` into a new headless coding agent package that can be reused across multiple UI implementations (TUI, VS Code extension, web interface).
+This document proposes extracting the agent infrastructure from `@mariozechner/pi-web-ui` into two new packages:
+
+1. **`@mariozechner/agent`** - General-purpose agent package with transport abstraction, state management, and attachment support
+2. **`@mariozechner/coding-agent`** - Specialized coding agent built on the general agent, with file manipulation tools and session management
 
 The new architecture will provide:
-- **Headless agent core** with file manipulation tools (read, bash, edit, write)
-- **Session management** for conversation persistence and resume capability
+- **General agent core** with transport abstraction (ProviderTransport, AppTransport)
+- **Reactive state management** with subscribe/emit pattern
+- **Attachment support** (type definitions only - processing stays in consumers)
+- **Message transformation** pipeline for filtering and adapting messages
+- **Message queueing** for out-of-band message injection
 - **Full abort support** throughout the execution pipeline
 - **Event-driven API** for flexible UI integration
 - **Clean separation** between agent logic and presentation layer
+- **Coding-specific tools** (read, bash, edit, write) in a specialized package
+- **Session management** for conversation persistence and resume capability
 
 ## Current Architecture Analysis
 
@@ -18,7 +26,7 @@ The new architecture will provide:
 ```
 pi-mono/
 ├── packages/ai/              # Core AI streaming (GOOD - keep as-is)
-├── packages/web-ui/          # Web UI with agent (GOOD - keep separate)
+├── packages/web-ui/          # Web UI with embedded agent (EXTRACT core agent logic)
 ├── packages/agent/           # OLD - needs to be replaced
 ├── packages/tui/             # Terminal UI lib (GOOD - low-level primitives)
 ├── packages/proxy/           # CORS proxy (unrelated)
@@ -77,9 +85,9 @@ interface AgentToolResult<T> {
 }
 ```
 
-### packages/web-ui/agent - Web Agent
+### packages/web-ui/src/agent - Web Agent
 
-**Status:** ✅ Good for web use cases, keep separate
+**Status:** ✅ KEEP AS-IS for now, will be replaced later after new packages are proven
 
 **Architecture:**
 ```typescript
@@ -94,60 +102,49 @@ class Agent {
   async prompt(input: string, attachments?: Attachment[]): Promise<void>
   abort(): void
   subscribe(fn: (e: AgentEvent) => void): () => void
+  setSystemPrompt(v: string): void
+  setModel(m: Model<any>): void
+  setThinkingLevel(l: ThinkingLevel): void
+  setTools(t: AgentTool<any>[]): void
+  replaceMessages(ms: AppMessage[]): void
+  appendMessage(m: AppMessage): void
+  async queueMessage(m: AppMessage): Promise<void>
+  clearMessages(): void
 }
 ```
 
-**Key Features:**
-- **Transport abstraction** (ProviderTransport for direct API, AppTransport for server-side)
-- **Attachment handling** (images, documents with text extraction)
-- **Message transformation** (app messages → LLM messages)
-- **Reactive state** (subscribe pattern for UI updates)
-- **Message queue** for injecting tool results/errors asynchronously
+**Key Features (will be basis for new `@mariozechner/agent` package):**
+- ✅ **Transport abstraction** (ProviderTransport for direct API, AppTransport for server-side proxy)
+- ✅ **Attachment type definition** (id, type, fileName, mimeType, size, content, extractedText, preview)
+- ✅ **Message transformation** pipeline (app messages → LLM messages, with filtering)
+- ✅ **Reactive state** (subscribe/emit pattern for UI updates)
+- ✅ **Message queueing** for injecting messages out-of-band during agent loop
+- ✅ **Abort support** (AbortController per prompt)
+- ✅ **State management** (systemPrompt, model, thinkingLevel, tools, messages, isStreaming, etc.)
 
-**Why it's different from coding agent:**
-- Browser-specific concerns (CORS, attachments)
-- Transport layer for flexible API routing
-- Tied to web UI state management
-- Supports rich media attachments
+**Strategy:**
+1. Use this implementation as the **reference design** for `@mariozechner/agent`
+2. Create new `@mariozechner/agent` package by copying/adapting this code
+3. Keep web-ui using its own embedded agent until new package is proven stable
+4. Eventually migrate web-ui to use `@mariozechner/agent` (Phase 2 of migration)
+5. Document processing (PDF/DOCX/PPTX/Excel) stays in web-ui permanently
 
 ### packages/agent - OLD Implementation
 
-**Status:** ⚠️ MUST BE REPLACED
+**Status:** ⚠️ REMOVE COMPLETELY
 
-**Architecture:**
-```typescript
-class Agent {
-  constructor(
-    config: AgentConfig,
-    renderer?: AgentEventReceiver,
-    sessionManager?: SessionManager
-  )
+**Why it should be removed:**
+1. **Tightly coupled to OpenAI SDK** - Not provider-agnostic, hardcoded to OpenAI's API
+2. **Outdated architecture** - Superseded by web-ui's better agent design
+3. **Mixed concerns** - Agent logic + tool implementations + rendering all in one package
+4. **Limited scope** - Cannot be reused across different UI implementations
 
-  async ask(userMessage: string): Promise<void>
-  interrupt(): void
-  setEvents(events: AgentEvent[]): void
-}
-```
+**What to salvage before removal:**
+1. **SessionManager** - Port to `@mariozechner/coding-agent` (JSONL-based session persistence)
+2. **Tool implementations** - Adapt read, bash, edit, write tools for coding-agent
+3. **Renderer abstractions** - Port TuiRenderer/ConsoleRenderer/JsonRenderer concepts to coding-agent-tui
 
-**Problems:**
-1. **Tightly coupled to OpenAI SDK** (not provider-agnostic)
-2. **Hardcoded tools** (read, list, bash, glob, rg)
-3. **Mixed concerns** (agent logic + tool implementations in same package)
-4. **No separation** between core loop and UI rendering
-5. **Two API paths** (completions vs responses) with branching logic
-
-**Good parts to preserve:**
-1. **SessionManager** - JSONL-based session persistence
-2. **Event receiver pattern** - Clean UI integration
-3. **Abort support** - Proper signal handling
-4. **Renderer abstraction** (ConsoleRenderer, TuiRenderer, JsonRenderer)
-
-**Tools implemented:**
-- `read`: Read file contents (1MB limit with truncation)
-- `list`: List directory contents
-- `bash`: Execute shell command with abort support
-- `glob`: Find files matching glob pattern
-- `rg`: Run ripgrep search
+**Action:** Delete this package entirely after extracting useful components to the new packages.
 
 ## Proposed Architecture
 
@@ -155,132 +152,440 @@ class Agent {
 
 ```
 pi-mono/
-├── packages/ai/                          # [unchanged] Core streaming
-├── packages/coding-agent/                # [NEW] Headless coding agent
+├── packages/ai/                          # [unchanged] Core streaming library
+│
+├── packages/agent/                       # [NEW] General-purpose agent
 │   ├── src/
-│   │   ├── agent.ts                      # Main agent class
-│   │   ├── session-manager.ts            # Session persistence
+│   │   ├── agent.ts                      # Main Agent class
+│   │   ├── types.ts                      # AgentState, AgentEvent, Attachment, etc.
+│   │   ├── transports/
+│   │   │   ├── types.ts                  # AgentTransport interface
+│   │   │   ├── ProviderTransport.ts      # Direct API calls
+│   │   │   ├── AppTransport.ts           # Server-side proxy
+│   │   │   ├── proxy-types.ts            # Proxy event types
+│   │   │   └── index.ts                  # Transport exports
+│   │   └── index.ts                      # Public API
+│   └── package.json
+│
+├── packages/coding-agent/                # [NEW] Coding-specific agent + CLI
+│   ├── src/
+│   │   ├── coding-agent.ts               # CodingAgent wrapper (uses @mariozechner/agent)
+│   │   ├── session-manager.ts            # Session persistence (JSONL)
 │   │   ├── tools/
 │   │   │   ├── read-tool.ts              # Read files (with pagination)
 │   │   │   ├── bash-tool.ts              # Shell execution
 │   │   │   ├── edit-tool.ts              # File editing (old_string → new_string)
 │   │   │   ├── write-tool.ts             # File creation/replacement
 │   │   │   └── index.ts                  # Tool exports
-│   │   └── types.ts                      # Public types
-│   └── package.json
+│   │   ├── cli/
+│   │   │   ├── index.ts                  # CLI entry point
+│   │   │   ├── renderers/
+│   │   │   │   ├── tui-renderer.ts       # Rich terminal UI
+│   │   │   │   ├── console-renderer.ts   # Simple console output
+│   │   │   │   └── json-renderer.ts      # JSONL output for piping
+│   │   │   └── main.ts                   # CLI app logic
+│   │   ├── types.ts                      # Public types
+│   │   └── index.ts                      # Public API (agent + tools)
+│   └── package.json                      # Exports both library + CLI binary
 │
-├── packages/coding-agent-tui/            # [NEW] Terminal interface
+├── packages/web-ui/                      # [updated] Uses @mariozechner/agent
 │   ├── src/
-│   │   ├── cli.ts                        # CLI entry point
-│   │   ├── renderers/
-│   │   │   ├── tui-renderer.ts           # Rich terminal UI
-│   │   │   ├── console-renderer.ts       # Simple console output
-│   │   │   └── json-renderer.ts          # JSONL output for piping
-│   │   └── main.ts                       # App logic
-│   └── package.json
+│   │   ├── utils/
+│   │   │   └── attachment-utils.ts       # Document processing (keep here)
+│   │   └── ...                           # Other web UI code
+│   └── package.json                      # Now depends on @mariozechner/agent
 │
-├── packages/web-ui/                      # [unchanged] Web UI keeps its own agent
 └── packages/tui/                         # [unchanged] Low-level terminal primitives
 ```
 
 ### Dependency Graph
 
 ```
-┌─────────────────────┐
-│   @mariozechner/    │
-│      pi-ai          │  ← Core streaming, tool interface
-└──────────┬──────────┘
+                    ┌─────────────────────┐
+                    │   @mariozechner/    │
+                    │      pi-ai          │  ← Core streaming, tool interface
+                    └──────────┬──────────┘
                                │ depends on
                                ↓
-┌─────────────────────┐
-│  @mariozechner/     │
-│   coding-agent      │  ← Headless agent + file tools
-└──────────┬──────────┘
+                    ┌─────────────────────┐
+                    │  @mariozechner/     │
+                    │      agent          │  ← General agent (transports, state, attachments)
+                    └──────────┬──────────┘
                                │ depends on
                                ↓
-    ┌──────────┬──────────┐
-    ↓          ↓          ↓
-┌────────┐ ┌───────┐ ┌────────┐
-│ TUI    │ │ VSCode│ │ Web UI │
-│ Client │ │  Ext  │ │ (own)  │
-└────────┘ └───────┘ └────────┘
+               ┌───────────────┴───────────────┐
+               ↓                               ↓
+    ┌─────────────────────┐         ┌─────────────────────┐
+    │  @mariozechner/     │         │  @mariozechner/     │
+    │   coding-agent      │         │     pi-web-ui       │
+    │ (lib + CLI + tools) │         │ (+ doc processing)  │
+    └─────────────────────┘         └─────────────────────┘
+```
+
+## Package: @mariozechner/agent
+
+### Core Types
+
+```typescript
+export interface Attachment {
+  id: string;
+  type: "image" | "document";
+  fileName: string;
+  mimeType: string;
+  size: number;
+  content: string;        // base64 encoded (without data URL prefix)
+  extractedText?: string; // For documents
+  preview?: string;       // base64 image preview
+}
+
+export type ThinkingLevel = "off" | "minimal" | "low" | "medium" | "high";
+
+// AppMessage abstraction - extends base LLM messages with app-specific features
+export type UserMessageWithAttachments = UserMessage & { attachments?: Attachment[] };
+
+// Extensible interface for custom app messages (via declaration merging)
+// Apps can add their own message types:
+// declare module "@mariozechner/agent" {
+//   interface CustomMessages {
+//     artifact: ArtifactMessage;
+//     notification: NotificationMessage;
+//   }
+// }
+export interface CustomMessages {
+  // Empty by default - apps extend via declaration merging
+}
+
+// AppMessage: Union of LLM messages + attachments + custom messages
+export type AppMessage =
+  | AssistantMessage
+  | UserMessageWithAttachments
+  | ToolResultMessage
+  | CustomMessages[keyof CustomMessages];
+
+export interface AgentState {
+  systemPrompt: string;
+  model: Model<any>;
+  thinkingLevel: ThinkingLevel;
+  tools: AgentTool<any>[];
+  messages: AppMessage[];      // Can include attachments + custom message types
+  isStreaming: boolean;
+  streamMessage: Message | null;
+  pendingToolCalls: Set<string>;
+  error?: string;
+}
+
+export type AgentEvent =
+  | { type: "state-update"; state: AgentState }
+  | { type: "started" }
+  | { type: "completed" };
+```
+
+### AppMessage Abstraction
+
+The `AppMessage` type is a key abstraction that extends base LLM messages with app-specific features while maintaining type safety and extensibility.
+
+**Key Benefits:**
+1. **Extends base messages** - Adds `attachments` field to `UserMessage` for file uploads
+2. **Type-safe extensibility** - Apps can add custom message types via declaration merging
+3. **Backward compatible** - Works seamlessly with base LLM messages from `@mariozechner/pi-ai`
+4. **Message transformation** - Filters app-specific fields before sending to LLM
+
+**Usage Example (Web UI):**
+```typescript
+import type { AppMessage } from "@mariozechner/agent";
+
+// Extend with custom message type for artifacts
+declare module "@mariozechner/agent" {
+  interface CustomMessages {
+    artifact: ArtifactMessage;
+  }
+}
+
+interface ArtifactMessage {
+  role: "artifact";
+  action: "create" | "update" | "delete";
+  filename: string;
+  content?: string;
+  title?: string;
+  timestamp: string;
+}
+
+// Now AppMessage includes: AssistantMessage | UserMessageWithAttachments | ToolResultMessage | ArtifactMessage
+const messages: AppMessage[] = [
+  { role: "user", content: "Hello", attachments: [attachment] },
+  { role: "assistant", content: [{ type: "text", text: "Hi!" }], /* ... */ },
+  { role: "artifact", action: "create", filename: "test.ts", content: "...", timestamp: "..." }
+];
+```
+
+**Usage Example (Coding Agent):**
+```typescript
+import type { AppMessage } from "@mariozechner/agent";
+
+// Coding agent can extend with session metadata
+declare module "@mariozechner/agent" {
+  interface CustomMessages {
+    session_metadata: SessionMetadataMessage;
+  }
+}
+
+interface SessionMetadataMessage {
+  role: "session_metadata";
+  sessionId: string;
+  timestamp: string;
+  workingDirectory: string;
+}
+```
+
+**Message Transformation:**
+
+The `messageTransformer` function converts app messages to LLM-compatible messages, including handling attachments:
+
+```typescript
+function defaultMessageTransformer(messages: AppMessage[]): Message[] {
+  return messages
+    .filter((m) => {
+      // Only keep standard LLM message roles
+      return m.role === "user" || m.role === "assistant" || m.role === "toolResult";
+    })
+    .map((m) => {
+      if (m.role === "user") {
+        const { attachments, ...baseMessage } = m as any;
+
+        // If no attachments, return as-is
+        if (!attachments || attachments.length === 0) {
+          return baseMessage as Message;
+        }
+
+        // Convert attachments to content blocks
+        const content = Array.isArray(baseMessage.content)
+          ? [...baseMessage.content]
+          : [{ type: "text", text: baseMessage.content }];
+
+        for (const attachment of attachments) {
+          // Add image blocks for image attachments
+          if (attachment.type === "image") {
+            content.push({
+              type: "image",
+              data: attachment.content,
+              mimeType: attachment.mimeType
+            });
+          }
+          // Add text blocks for documents with extracted text
+          else if (attachment.type === "document" && attachment.extractedText) {
+            content.push({
+              type: "text",
+              text: attachment.extractedText
+            });
+          }
+        }
+
+        return { ...baseMessage, content } as Message;
+      }
+      return m as Message;
+    });
+}
+```
+
+This ensures that:
+- Custom message types (like `artifact`, `session_metadata`) are filtered out
+- Image attachments are converted to `ImageContent` blocks
+- Document attachments with extracted text are converted to `TextContent` blocks
+- The `attachments` field itself is stripped (replaced by proper content blocks)
+- LLM receives only standard `Message` types from `@mariozechner/pi-ai`
+
+### Agent Class
+
+```typescript
+export interface AgentOptions {
+  initialState?: Partial<AgentState>;
+  transport: AgentTransport;
+  // Transform app messages to LLM-compatible messages before sending
+  messageTransformer?: (messages: AppMessage[]) => Message[] | Promise<Message[]>;
+}
+
+export class Agent {
+  constructor(opts: AgentOptions);
+
+  get state(): AgentState;
+  subscribe(fn: (e: AgentEvent) => void): () => void;
+
+  // State mutators
+  setSystemPrompt(v: string): void;
+  setModel(m: Model<any>): void;
+  setThinkingLevel(l: ThinkingLevel): void;
+  setTools(t: AgentTool<any>[]): void;
+  replaceMessages(ms: AppMessage[]): void;
+  appendMessage(m: AppMessage): void;
+  async queueMessage(m: AppMessage): Promise<void>;
+  clearMessages(): void;
+
+  // Main prompt method
+  async prompt(input: string, attachments?: Attachment[]): Promise<void>;
+
+  // Abort current operation
+  abort(): void;
+}
+```
+
+**Key Features:**
+1. **Reactive state** - Subscribe to state updates for UI binding
+2. **Transport abstraction** - Pluggable backends (direct API, proxy server, etc.)
+3. **Message transformation** - Convert app-specific messages to LLM format
+4. **Message queueing** - Inject messages during agent loop (for tool results, errors)
+5. **Attachment support** - Type-safe attachment handling (processing is external)
+6. **Abort support** - Cancel in-progress operations
+
+### Transport Interface
+
+```typescript
+export interface AgentRunConfig {
+  systemPrompt: string;
+  tools: AgentTool<any>[];
+  model: Model<any>;
+  reasoning?: "low" | "medium" | "high";
+  getQueuedMessages?: <T>() => Promise<QueuedMessage<T>[]>;
+}
+
+export interface AgentTransport {
+  run(
+    messages: Message[],
+    userMessage: Message,
+    config: AgentRunConfig,
+    signal?: AbortSignal,
+  ): AsyncIterable<AgentEvent>;
+}
+```
+
+### ProviderTransport
+
+```typescript
+export class ProviderTransport implements AgentTransport {
+  async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
+    // Calls LLM providers directly using agentLoop from @mariozechner/pi-ai
+    // Optionally routes through CORS proxy if configured
+  }
+}
+```
+
+### AppTransport
+
+```typescript
+export class AppTransport implements AgentTransport {
+  constructor(proxyUrl: string);
+
+  async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
+    // Routes requests through app server with user authentication
+    // Server manages API keys and usage tracking
+  }
+}
 ```
 
 ## Package: @mariozechner/coding-agent
 
-### Core Agent Class
+### CodingAgent Class
 
 ```typescript
-export interface CodingAgentConfig {
+export interface CodingAgentOptions {
   systemPrompt: string;
   model: Model<any>;
   reasoning?: "low" | "medium" | "high";
   apiKey: string;
-}
-
-export interface CodingAgentOptions {
-  config: CodingAgentConfig;
-  sessionManager?: SessionManager;
   workingDirectory?: string;
+  sessionManager?: SessionManager;
 }
 
 export class CodingAgent {
   constructor(options: CodingAgentOptions);
 
+  // Access underlying agent
+  get agent(): Agent;
+
+  // State accessors
+  get state(): AgentState;
+  subscribe(fn: (e: AgentEvent) => void): () => void;
+
   // Send a message to the agent
-  async prompt(message: string, signal?: AbortSignal): AsyncIterable<AgentEvent>;
+  async prompt(message: string, attachments?: Attachment[]): Promise<void>;
 
-  // Restore from session events (for --continue mode)
-  setMessages(messages: Message[]): void;
+  // Abort current operation
+  abort(): void;
 
-  // Get current message history
-  getMessages(): Message[];
+  // Message management for session restoration
+  replaceMessages(messages: AppMessage[]): void;
+  getMessages(): AppMessage[];
 }
 ```
 
 **Key design decisions:**
-1. **AsyncIterable instead of callbacks** - More flexible for consumers
-2. **Signal per prompt** - Each prompt() call accepts its own AbortSignal
-3. **No internal state management** - Consumers handle UI state
-4. **Simple message management** - Get/set for session restoration
+1. **Wraps @mariozechner/agent** - Builds on the general agent package
+2. **Pre-configured tools** - Includes read, bash, edit, write tools
+3. **Session management** - Optional JSONL-based session persistence
+4. **Working directory context** - All file operations relative to this directory
+5. **Simple API** - Hides transport complexity, uses ProviderTransport by default
 
 ### Usage Example (TUI)
 
 ```typescript
 import { CodingAgent } from "@mariozechner/coding-agent";
 import { SessionManager } from "@mariozechner/coding-agent";
+import { getModel } from "@mariozechner/pi-ai";
 
 const session = new SessionManager({ continue: true });
 const agent = new CodingAgent({
-  config: {
   systemPrompt: "You are a coding assistant...",
   model: getModel("openai", "gpt-4"),
   apiKey: process.env.OPENAI_API_KEY!,
-  },
-  sessionManager: session,
   workingDirectory: process.cwd(),
+  sessionManager: session,
 });
 
 // Restore previous session
 if (session.hasData()) {
-  agent.setMessages(session.getMessages());
+  agent.replaceMessages(session.getMessages());
 }
 
-// Send prompt with abort support
-const controller = new AbortController();
-for await (const event of agent.prompt("Fix the bug in server.ts", controller.signal)) {
-  switch (event.type) {
-    case "message_update":
-      renderer.updateAssistant(event.message);
-      break;
-    case "tool_execution_start":
-      renderer.showTool(event.toolName, event.args);
-      break;
-    case "tool_execution_end":
-      renderer.showToolResult(event.toolName, event.result);
-      break;
+// Subscribe to state changes
+agent.subscribe((event) => {
+  if (event.type === "state-update") {
+    renderer.render(event.state);
+  } else if (event.type === "completed") {
+    session.save(agent.getMessages());
   }
-}
+});
+
+// Send prompt
+await agent.prompt("Fix the bug in server.ts");
+```
+
+### Usage Example (Web UI)
+
+```typescript
+import { Agent, ProviderTransport, Attachment } from "@mariozechner/agent";
+import { getModel } from "@mariozechner/pi-ai";
+import { loadAttachment } from "./utils/attachment-utils"; // Web UI keeps this
+
+const agent = new Agent({
+  transport: new ProviderTransport(),
+  initialState: {
+    systemPrompt: "You are a helpful assistant...",
+    model: getModel("google", "gemini-2.5-flash"),
+    thinkingLevel: "low",
+    tools: [],
+  },
+});
+
+// Subscribe to state changes for UI updates
+agent.subscribe((event) => {
+  if (event.type === "state-update") {
+    updateUI(event.state);
+  }
+});
+
+// Handle file upload and send prompt
+const file = await fileInput.files[0];
+const attachment = await loadAttachment(file); // Processes PDF/DOCX/etc
+await agent.prompt("Analyze this document", [attachment]);
 ```
 
 ### Session Manager
@@ -300,8 +605,7 @@ export interface SessionMetadata {
 
 export interface SessionData {
   metadata: SessionMetadata;
-  messages: Message[];          // Conversation history
-  totalUsage: TokenUsage;       // Aggregated token usage
+  messages: AppMessage[];       // Conversation history
 }
 
 export class SessionManager {
@@ -310,8 +614,8 @@ export class SessionManager {
   // Start a new session (writes metadata)
   startSession(config: CodingAgentConfig): void;
 
-  // Log an event (appends to JSONL)
-  appendEvent(event: AgentEvent): void;
+  // Append a message to the session (appends to JSONL)
+  appendMessage(message: AppMessage): void;
 
   // Check if session has existing data
   hasData(): boolean;
@@ -320,7 +624,7 @@ export class SessionManager {
   getData(): SessionData | null;
 
   // Get just the messages for agent restoration
-  getMessages(): Message[];
+  getMessages(): AppMessage[];
 
   // Get session file path
   getFilePath(): string;
@@ -332,12 +636,19 @@ export class SessionManager {
 
 **Session Storage Format (JSONL):**
 ```jsonl
-{"type":"session","id":"uuid","timestamp":"2025-10-12T10:00:00Z","cwd":"/path","config":{...}}
-{"type":"event","timestamp":"2025-10-12T10:00:01Z","event":{"type":"turn_start"}}
-{"type":"event","timestamp":"2025-10-12T10:00:02Z","event":{"type":"message_start",...}}
-{"type":"event","timestamp":"2025-10-12T10:00:03Z","event":{"type":"message_end",...}}
+{"type":"metadata","id":"uuid","timestamp":"2025-10-12T10:00:00Z","cwd":"/path","config":{...}}
+{"type":"message","message":{"role":"user","content":"Fix the bug in server.ts"}}
+{"type":"message","message":{"role":"assistant","content":[{"type":"text","text":"I'll help..."}],...}}
+{"type":"message","message":{"role":"toolResult","toolCallId":"call_123","output":"..."}}
+{"type":"message","message":{"role":"assistant","content":[{"type":"text","text":"Fixed!"}],...}}
 ```
 
+**How it works:**
+- First line is session metadata (id, timestamp, working directory, config)
+- Each subsequent line is an `AppMessage` from `agent.state.messages`
+- Messages are appended as they're added to the agent state (append-only)
+- On session restore, read all message lines to reconstruct conversation history
+
 **Session File Naming:**
 ```
 ~/.pi/sessions/--path-to-project--/
@@ -643,9 +954,11 @@ export class WriteTool implements AgentTool<typeof writeToolSchema, WriteToolDet
 }
 ```
 
-## Package: @mariozechner/coding-agent-tui
+## CLI Interface (included in @mariozechner/coding-agent)
 
-### CLI Interface
+The coding-agent package includes both a library and a CLI interface in one package.
+
+### CLI Usage
 
 ```bash
 # Interactive mode (default)
@@ -667,7 +980,7 @@ coding-agent --model openai/gpt-4 --api-key $KEY
 coding-agent --json < prompts.jsonl > results.jsonl
 ```
 
-### Arguments
+### CLI Arguments
 
 ```typescript
 {
@@ -818,33 +1131,58 @@ app.listen(3000);
 
 ## Migration Plan
 
-### Phase 1: Extract Core Package
+### Phase 1: Create General Agent Package
+1. Create `packages/agent/` structure
+2. **COPY** Agent class from web-ui/src/agent/agent.ts (don't extract yet)
+3. Copy types (AgentState, AgentEvent, Attachment, DebugLogEntry, ThinkingLevel)
+4. Copy transports (types.ts, ProviderTransport.ts, AppTransport.ts, proxy-types.ts)
+5. Adapt code to work as standalone package
+6. Write unit tests for Agent class
+7. Write tests for both transports
+8. Publish `@mariozechner/agent@0.1.0`
+9. **Keep web-ui unchanged** - it continues using its embedded agent
+
+### Phase 2: Create Coding Agent Package (with CLI)
 1. Create `packages/coding-agent/` structure
 2. Port SessionManager from old agent package
-3. Implement BashTool, EditTool, WriteTool
-4. Implement CodingAgent class using pi-ai/agentLoop
-5. Write tests for each tool
-6. Write integration tests
+3. Implement ReadTool, BashTool, EditTool, WriteTool
+4. Implement CodingAgent class (wraps @mariozechner/agent)
+5. Implement CLI in `src/cli/` directory:
+   - CLI entry point (index.ts)
+   - TuiRenderer, ConsoleRenderer, JsonRenderer
+   - Argument parsing
+   - Interactive and single-shot modes
+6. Write tests for tools and agent
+7. Write integration tests for CLI
+8. Publish `@mariozechner/coding-agent@0.1.0` (includes library + CLI binary)
 
-### Phase 2: Build TUI
-1. Create `packages/coding-agent-tui/`
-2. Port TuiRenderer from old agent package
-3. Port ConsoleRenderer, JsonRenderer
-4. Implement CLI argument parsing
-5. Implement interactive and single-shot modes
-6. Test session resume functionality
+### Phase 3: Prove Out New Packages
+1. Use coding-agent (library + CLI) extensively
+2. Fix bugs and iterate on API design
+3. Gather feedback from real usage
+4. Ensure stability and performance
 
-### Phase 3: Update Dependencies
-1. Update web-ui if needed (should be unaffected)
-2. Deprecate old agent package
-3. Update documentation
-4. Update examples
+### Phase 4: Migrate Web UI (OPTIONAL, later)
+1. Once new `@mariozechner/agent` is proven stable
+2. Update web-ui package.json to depend on `@mariozechner/agent`
+3. Remove src/agent/agent.ts, src/agent/types.ts, src/agent/transports/
+4. Keep src/utils/attachment-utils.ts (document processing)
+5. Update imports to use `@mariozechner/agent`
+6. Test that web UI still works correctly
+7. Verify document attachments (PDF, DOCX, etc.) still work
 
-### Phase 4: Future Enhancements
-1. Build VS Code extension
-2. Add more tools (grep, find, etc.) as optional
+### Phase 5: Cleanup
+1. Deprecate/remove old `packages/agent/` package
+2. Update all documentation
+3. Create migration guide
+4. Add examples for all use cases
+
+### Phase 6: Future Enhancements
+1. Build VS Code extension using `@mariozechner/coding-agent`
+2. Add more tools (grep, find, glob, etc.) as optional plugins
 3. Plugin system for custom tools
 4. Parallel tool execution
+5. Streaming tool output for long-running commands
 
 ## Open Questions & Decisions
 
@@ -975,14 +1313,29 @@ Error: Cannot read binary file 'dist/app.js'. Use bash tool if you need to inspe
 
 This architecture provides:
 
-✅ **Headless core** - Clean separation between agent logic and UI
-✅ **Reusable** - Same agent for TUI, VS Code, web, APIs
-✅ **Composable** - Build on pi-ai primitives
-✅ **Abortable** - First-class cancellation support
-✅ **Session persistence** - Resume conversations seamlessly
+### General Agent Package (`@mariozechner/agent`)
+✅ **Transport abstraction** - Pluggable backends (ProviderTransport, AppTransport)
+✅ **Reactive state** - Subscribe/emit pattern for UI binding
+✅ **Message transformation** - Flexible pipeline for message filtering/adaptation
+✅ **Message queueing** - Out-of-band message injection during agent loop
+✅ **Attachment support** - Type-safe attachment handling (processing is external)
+✅ **Abort support** - First-class cancellation with AbortController
+✅ **Provider agnostic** - Works with any LLM provider via @mariozechner/pi-ai
+✅ **Type-safe** - Full TypeScript with proper types
+
+### Coding Agent Package (`@mariozechner/coding-agent`)
+✅ **Builds on general agent** - Leverages transport abstraction and state management
+✅ **Session persistence** - JSONL-based session storage and resume
 ✅ **Focused tools** - read, bash, edit, write (4 tools, no more)
-✅ **Smart pagination** - 5000-line chunks with offset/limit
-✅ **Type-safe** - Full TypeScript with schema validation
+✅ **Smart pagination** - 5000-line chunks with offset/limit for ReadTool
+✅ **Working directory context** - All tools operate relative to project root
+✅ **Simple API** - Hides complexity, easy to use
 ✅ **Testable** - Pure functions, mockable dependencies
 
-The key insight is to **keep web-ui's agent separate** (it has different concerns) while creating a **new focused coding agent** for file manipulation workflows that can be shared across non-web interfaces.
+### Key Architectural Insights
+1. **Extract, don't rewrite** - The web-ui agent is well-designed; extract it into a general package
+2. **Separation of concerns** - Document processing (PDF/DOCX/etc.) stays in web-ui, only type definitions move to general agent
+3. **Layered architecture** - pi-ai → agent → coding-agent → coding-agent-tui
+4. **Reusable across UIs** - Web UI and coding agent both use the same general agent package
+5. **Pluggable transports** - Easy to add new backends (local API, proxy server, etc.)
+6. **Attachment flexibility** - Type is defined centrally, processing is done by consumers

package-lock.json

@@ -728,6 +728,123 @@
 				"hono": "^4"
 			}
 		},
+		"node_modules/@isaacs/balanced-match": {
+			"version": "4.0.1",
+			"resolved": "https://registry.npmjs.org/@isaacs/balanced-match/-/balanced-match-4.0.1.tgz",
+			"integrity": "sha512-yzMTt9lEb8Gv7zRioUilSglI0c0smZ9k5D65677DLWLtWJaXIS3CqcGyUFByYKlnUj6TkjLVs54fBl6+TiGQDQ==",
+			"license": "MIT",
+			"engines": {
+				"node": "20 || >=22"
+			}
+		},
+		"node_modules/@isaacs/brace-expansion": {
+			"version": "5.0.0",
+			"resolved": "https://registry.npmjs.org/@isaacs/brace-expansion/-/brace-expansion-5.0.0.tgz",
+			"integrity": "sha512-ZT55BDLV0yv0RBm2czMiZ+SqCGO7AvmOM3G/w2xhVPH+te0aKgFjmBvGlL1dH+ql2tgGO3MVrbb3jCKyvpgnxA==",
+			"license": "MIT",
+			"dependencies": {
+				"@isaacs/balanced-match": "^4.0.1"
+			},
+			"engines": {
+				"node": "20 || >=22"
+			}
+		},
+		"node_modules/@isaacs/cliui": {
+			"version": "8.0.2",
+			"resolved": "https://registry.npmjs.org/@isaacs/cliui/-/cliui-8.0.2.tgz",
+			"integrity": "sha512-O8jcjabXaleOG9DQ0+ARXWZBTfnP4WNAqzuiJK7ll44AmxGKv/J2M4TPjxjY3znBCfvBXFzucm1twdyFybFqEA==",
+			"license": "ISC",
+			"dependencies": {
+				"string-width": "^5.1.2",
+				"string-width-cjs": "npm:string-width@^4.2.0",
+				"strip-ansi": "^7.0.1",
+				"strip-ansi-cjs": "npm:strip-ansi@^6.0.1",
+				"wrap-ansi": "^8.1.0",
+				"wrap-ansi-cjs": "npm:wrap-ansi@^7.0.0"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/@isaacs/cliui/node_modules/ansi-regex": {
+			"version": "6.2.2",
+			"resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.2.2.tgz",
+			"integrity": "sha512-Bq3SmSpyFHaWjPk8If9yc6svM8c56dB5BAtW4Qbw5jHTwwXXcTLoRMkpDJp6VL0XzlWaCHTXrkFURMYmD0sLqg==",
+			"license": "MIT",
+			"engines": {
+				"node": ">=12"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/ansi-regex?sponsor=1"
+			}
+		},
+		"node_modules/@isaacs/cliui/node_modules/ansi-styles": {
+			"version": "6.2.3",
+			"resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-6.2.3.tgz",
+			"integrity": "sha512-4Dj6M28JB+oAH8kFkTLUo+a2jwOFkuqb3yucU0CANcRRUbxS0cP0nZYCGjcc3BNXwRIsUVmDGgzawme7zvJHvg==",
+			"license": "MIT",
+			"engines": {
+				"node": ">=12"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/ansi-styles?sponsor=1"
+			}
+		},
+		"node_modules/@isaacs/cliui/node_modules/emoji-regex": {
+			"version": "9.2.2",
+			"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz",
+			"integrity": "sha512-L18DaJsXSUk2+42pv8mLs5jJT2hqFkFE4j21wOmgbUqsZ2hL72NsUU785g9RXgo3s0ZNgVl42TiHp3ZtOv/Vyg==",
+			"license": "MIT"
+		},
+		"node_modules/@isaacs/cliui/node_modules/string-width": {
+			"version": "5.1.2",
+			"resolved": "https://registry.npmjs.org/string-width/-/string-width-5.1.2.tgz",
+			"integrity": "sha512-HnLOCR3vjcY8beoNLtcjZ5/nxn2afmME6lhrDrebokqMap+XbeW8n9TXpPDOqdGK5qcI3oT0GKTW6wC7EMiVqA==",
+			"license": "MIT",
+			"dependencies": {
+				"eastasianwidth": "^0.2.0",
+				"emoji-regex": "^9.2.2",
+				"strip-ansi": "^7.0.1"
+			},
+			"engines": {
+				"node": ">=12"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/sindresorhus"
+			}
+		},
+		"node_modules/@isaacs/cliui/node_modules/strip-ansi": {
+			"version": "7.1.2",
+			"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.1.2.tgz",
+			"integrity": "sha512-gmBGslpoQJtgnMAvOVqGZpEz9dyoKTCzy2nfz/n8aIFhN/jCE/rCmcxabB6jOOHV+0WNnylOxaxBQPSvcWklhA==",
+			"license": "MIT",
+			"dependencies": {
+				"ansi-regex": "^6.0.1"
+			},
+			"engines": {
+				"node": ">=12"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/strip-ansi?sponsor=1"
+			}
+		},
+		"node_modules/@isaacs/cliui/node_modules/wrap-ansi": {
+			"version": "8.1.0",
+			"resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-8.1.0.tgz",
+			"integrity": "sha512-si7QWI6zUMq56bESFvagtmzMdGOtoxfR+Sez11Mobfc7tm+VkUckk9bW2UeffTGVUbOksxmSw0AA2gs8g71NCQ==",
+			"license": "MIT",
+			"dependencies": {
+				"ansi-styles": "^6.1.0",
+				"string-width": "^5.0.1",
+				"strip-ansi": "^7.0.1"
+			},
+			"engines": {
+				"node": ">=12"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/wrap-ansi?sponsor=1"
+			}
+		},
 		"node_modules/@isaacs/fs-minipass": {
 			"version": "4.0.1",
 			"resolved": "https://registry.npmjs.org/@isaacs/fs-minipass/-/fs-minipass-4.0.1.tgz",
@@ -806,6 +923,10 @@
 				"@lit-labs/ssr-dom-shim": "^1.4.0"
 			}
 		},
+		"node_modules/@mariozechner/coding-agent": {
+			"resolved": "packages/coding-agent",
+			"link": true
+		},
 		"node_modules/@mariozechner/jailjs": {
 			"version": "0.1.1",
 			"resolved": "https://registry.npmjs.org/@mariozechner/jailjs/-/jailjs-0.1.1.tgz",
@@ -845,6 +966,10 @@
 			"resolved": "packages/agent",
 			"link": true
 		},
+		"node_modules/@mariozechner/pi-agent-old": {
+			"resolved": "packages/agent-old",
+			"link": true
+		},
 		"node_modules/@mariozechner/pi-ai": {
 			"resolved": "packages/ai",
 			"link": true
@@ -1997,6 +2122,22 @@
 			"dev": true,
 			"license": "MIT"
 		},
+		"node_modules/@types/glob": {
+			"version": "8.1.0",
+			"resolved": "https://registry.npmjs.org/@types/glob/-/glob-8.1.0.tgz",
+			"integrity": "sha512-IO+MJPVhoqz+28h1qLAcBEH2+xHMK6MTyHJc7MTnnYb6wsoLR29POVGJ7LycmVXIqyy/4/2ShP5sUwTXuOwb/w==",
+			"license": "MIT",
+			"dependencies": {
+				"@types/minimatch": "^5.1.2",
+				"@types/node": "*"
+			}
+		},
+		"node_modules/@types/minimatch": {
+			"version": "5.1.2",
+			"resolved": "https://registry.npmjs.org/@types/minimatch/-/minimatch-5.1.2.tgz",
+			"integrity": "sha512-K0VQKziLUWkVKiRVrx4a40iPaxTUefQmjtkQofBkYRcoaaL/8rhwDWww9qWbrgicNOgnpIsMxyNIUM4+n6dUIA==",
+			"license": "MIT"
+		},
 		"node_modules/@types/node": {
 			"version": "22.18.8",
 			"resolved": "https://registry.npmjs.org/@types/node/-/node-22.18.8.tgz",
@@ -2543,6 +2684,20 @@
 				"node": ">=0.8"
 			}
 		},
+		"node_modules/cross-spawn": {
+			"version": "7.0.6",
+			"resolved": "https://registry.npmjs.org/cross-spawn/-/cross-spawn-7.0.6.tgz",
+			"integrity": "sha512-uV2QOWP2nWzsy2aMp8aRibhi9dlzF5Hgh5SHaB9OiTGEyDTiJJyx0uy51QXdyWbtAHNua4XJzUKca3OzKUd3vA==",
+			"license": "MIT",
+			"dependencies": {
+				"path-key": "^3.1.0",
+				"shebang-command": "^2.0.0",
+				"which": "^2.0.1"
+			},
+			"engines": {
+				"node": ">= 8"
+			}
+		},
 		"node_modules/debug": {
 			"version": "4.4.3",
 			"resolved": "https://registry.npmjs.org/debug/-/debug-4.4.3.tgz",
@@ -2624,6 +2779,12 @@
 				"jszip": ">=3.0.0"
 			}
 		},
+		"node_modules/eastasianwidth": {
+			"version": "0.2.0",
+			"resolved": "https://registry.npmjs.org/eastasianwidth/-/eastasianwidth-0.2.0.tgz",
+			"integrity": "sha512-I88TYZWc9XiYHRQ4/3c5rjjfgkjhLyW2luGIheGERbNQ6OY7yTybanSpDXZa8y7VUP9YmDcYa+eyq4ca7iLqWA==",
+			"license": "MIT"
+		},
 		"node_modules/ecdsa-sig-formatter": {
 			"version": "1.0.11",
 			"resolved": "https://registry.npmjs.org/ecdsa-sig-formatter/-/ecdsa-sig-formatter-1.0.11.tgz",
@@ -2637,7 +2798,6 @@
 			"version": "8.0.0",
 			"resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
 			"integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
-			"dev": true,
 			"license": "MIT"
 		},
 		"node_modules/end-of-stream": {
@@ -2812,6 +2972,22 @@
 				"node": ">=8"
 			}
 		},
+		"node_modules/foreground-child": {
+			"version": "3.3.1",
+			"resolved": "https://registry.npmjs.org/foreground-child/-/foreground-child-3.3.1.tgz",
+			"integrity": "sha512-gIXjKqtFuWEgzFRJA9WCQeSJLZDjgJUOMCMzxtvFq/37KojM1BFGufqsCy0r4qSQmYLsZYMeyRqzIWOMup03sw==",
+			"license": "ISC",
+			"dependencies": {
+				"cross-spawn": "^7.0.6",
+				"signal-exit": "^4.0.1"
+			},
+			"engines": {
+				"node": ">=14"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/isaacs"
+			}
+		},
 		"node_modules/frac": {
 			"version": "1.1.2",
 			"resolved": "https://registry.npmjs.org/frac/-/frac-1.1.2.tgz",
@@ -2903,6 +3079,29 @@
 			"dev": true,
 			"license": "MIT"
 		},
+		"node_modules/glob": {
+			"version": "11.0.3",
+			"resolved": "https://registry.npmjs.org/glob/-/glob-11.0.3.tgz",
+			"integrity": "sha512-2Nim7dha1KVkaiF4q6Dj+ngPPMdfvLJEOpZk/jKiUAkqKebpGAWQXAq9z1xu9HKu5lWfqw/FASuccEjyznjPaA==",
+			"license": "ISC",
+			"dependencies": {
+				"foreground-child": "^3.3.1",
+				"jackspeak": "^4.1.1",
+				"minimatch": "^10.0.3",
+				"minipass": "^7.1.2",
+				"package-json-from-dist": "^1.0.0",
+				"path-scurry": "^2.0.0"
+			},
+			"bin": {
+				"glob": "dist/esm/bin.mjs"
+			},
+			"engines": {
+				"node": "20 || >=22"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/isaacs"
+			}
+		},
 		"node_modules/google-auth-library": {
 			"version": "9.15.1",
 			"resolved": "https://registry.npmjs.org/google-auth-library/-/google-auth-library-9.15.1.tgz",
@@ -3112,6 +3311,27 @@
 			"integrity": "sha512-VLghIWNM6ELQzo7zwmcg0NmTVyWKYjvIeM83yjp0wRDTmUnrM678fQbcKBo6n2CJEF0szoG//ytg+TKla89ALQ==",
 			"license": "MIT"
 		},
+		"node_modules/isexe": {
+			"version": "2.0.0",
+			"resolved": "https://registry.npmjs.org/isexe/-/isexe-2.0.0.tgz",
+			"integrity": "sha512-RHxMLp9lnKHGHRng9QFhRCMbYAcVpn69smSGcq3f36xjgVVWThj4qqLbTLlq7Ssj8B+fIQ1EuCEGI2lKsyQeIw==",
+			"license": "ISC"
+		},
+		"node_modules/jackspeak": {
+			"version": "4.1.1",
+			"resolved": "https://registry.npmjs.org/jackspeak/-/jackspeak-4.1.1.tgz",
+			"integrity": "sha512-zptv57P3GpL+O0I7VdMJNBZCu+BPHVQUk55Ft8/QCJjTVxrnJHuVuX/0Bl2A6/+2oyR/ZMEuFKwmzqqZ/U5nPQ==",
+			"license": "BlueOak-1.0.0",
+			"dependencies": {
+				"@isaacs/cliui": "^8.0.2"
+			},
+			"engines": {
+				"node": "20 || >=22"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/isaacs"
+			}
+		},
 		"node_modules/jiti": {
 			"version": "2.6.1",
 			"resolved": "https://registry.npmjs.org/jiti/-/jiti-2.6.1.tgz",
@@ -3509,6 +3729,15 @@
 			"dev": true,
 			"license": "MIT"
 		},
+		"node_modules/lru-cache": {
+			"version": "11.2.2",
+			"resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-11.2.2.tgz",
+			"integrity": "sha512-F9ODfyqML2coTIsQpSkRHnLSZMtkU8Q+mSfcaIyKwy58u+8k5nvAYeiNhsyMARvzNcXJ9QfWVrcPsC9e9rAxtg==",
+			"license": "ISC",
+			"engines": {
+				"node": "20 || >=22"
+			}
+		},
 		"node_modules/lucide": {
 			"version": "0.544.0",
 			"resolved": "https://registry.npmjs.org/lucide/-/lucide-0.544.0.tgz",
@@ -3577,6 +3806,21 @@
 				"url": "https://github.com/sponsors/sindresorhus"
 			}
 		},
+		"node_modules/minimatch": {
+			"version": "10.0.3",
+			"resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.0.3.tgz",
+			"integrity": "sha512-IPZ167aShDZZUMdRk66cyQAW3qr0WzbHkPdMYa8bzZhlHhO3jALbKdxcaak7W9FfT2rZNpQuUu4Od7ILEpXSaw==",
+			"license": "ISC",
+			"dependencies": {
+				"@isaacs/brace-expansion": "^5.0.0"
+			},
+			"engines": {
+				"node": "20 || >=22"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/isaacs"
+			}
+		},
 		"node_modules/minimist": {
 			"version": "1.2.8",
 			"resolved": "https://registry.npmjs.org/minimist/-/minimist-1.2.8.tgz",
@@ -3738,6 +3982,12 @@
 				}
 			}
 		},
+		"node_modules/package-json-from-dist": {
+			"version": "1.0.1",
+			"resolved": "https://registry.npmjs.org/package-json-from-dist/-/package-json-from-dist-1.0.1.tgz",
+			"integrity": "sha512-UEZIS3/by4OC8vL3P2dTXRETpebLI2NiI5vIrjaD/5UtrkFX/tNbwjTSRAGC/+7CAo2pIcBaRgWmcBBHcsaCIw==",
+			"license": "BlueOak-1.0.0"
+		},
 		"node_modules/pako": {
 			"version": "1.0.11",
 			"resolved": "https://registry.npmjs.org/pako/-/pako-1.0.11.tgz",
@@ -3750,6 +4000,31 @@
 			"integrity": "sha512-Njv/59hHaokb/hRUjce3Hdv12wd60MtM9Z5Olmn+nehe0QDAsRtRbJPvJ0Z91TusF0SuZRIvnM+S4l6EIP8leA==",
 			"license": "MIT"
 		},
+		"node_modules/path-key": {
+			"version": "3.1.1",
+			"resolved": "https://registry.npmjs.org/path-key/-/path-key-3.1.1.tgz",
+			"integrity": "sha512-ojmeN0qd+y0jszEtoY48r0Peq5dwMEkIlCOu6Q5f41lfkswXuKtYrhgoTpLnyIcHm24Uhqx+5Tqm2InSwLhE6Q==",
+			"license": "MIT",
+			"engines": {
+				"node": ">=8"
+			}
+		},
+		"node_modules/path-scurry": {
+			"version": "2.0.0",
+			"resolved": "https://registry.npmjs.org/path-scurry/-/path-scurry-2.0.0.tgz",
+			"integrity": "sha512-ypGJsmGtdXUOeM5u93TyeIEfEhM6s+ljAhrk5vAvSx8uyY/02OvrZnA0YNGUrPXfpJMgI1ODd3nwz8Npx4O4cg==",
+			"license": "BlueOak-1.0.0",
+			"dependencies": {
+				"lru-cache": "^11.0.0",
+				"minipass": "^7.1.2"
+			},
+			"engines": {
+				"node": "20 || >=22"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/isaacs"
+			}
+		},
 		"node_modules/pathe": {
 			"version": "2.0.3",
 			"resolved": "https://registry.npmjs.org/pathe/-/pathe-2.0.3.tgz",
@@ -4023,6 +4298,27 @@
 			"integrity": "sha512-MATJdZp8sLqDl/68LfQmbP8zKPLQNV6BIZoIgrscFDQ+RsvK/BxeDQOgyxKKoh0y/8h3BqVFnCqQ/gd+reiIXA==",
 			"license": "MIT"
 		},
+		"node_modules/shebang-command": {
+			"version": "2.0.0",
+			"resolved": "https://registry.npmjs.org/shebang-command/-/shebang-command-2.0.0.tgz",
+			"integrity": "sha512-kHxr2zZpYtdmrN1qDjrrX/Z1rR1kG8Dx+gkpK1G4eXmvXswmcE1hTWBWYUzlraYw1/yZp6YuDY77YtvbN0dmDA==",
+			"license": "MIT",
+			"dependencies": {
+				"shebang-regex": "^3.0.0"
+			},
+			"engines": {
+				"node": ">=8"
+			}
+		},
+		"node_modules/shebang-regex": {
+			"version": "3.0.0",
+			"resolved": "https://registry.npmjs.org/shebang-regex/-/shebang-regex-3.0.0.tgz",
+			"integrity": "sha512-7++dFhtcx3353uBaq8DDR4NuxBetBzC7ZQOhmTQInHEd6bSrXdiEyzCvG07Z44UYdLShWUyXt5M/yhz8ekcb1A==",
+			"license": "MIT",
+			"engines": {
+				"node": ">=8"
+			}
+		},
 		"node_modules/shell-quote": {
 			"version": "1.8.3",
 			"resolved": "https://registry.npmjs.org/shell-quote/-/shell-quote-1.8.3.tgz",
@@ -4043,6 +4339,18 @@
 			"dev": true,
 			"license": "ISC"
 		},
+		"node_modules/signal-exit": {
+			"version": "4.1.0",
+			"resolved": "https://registry.npmjs.org/signal-exit/-/signal-exit-4.1.0.tgz",
+			"integrity": "sha512-bzyZ1e88w9O1iNJbKnOlvYTrWPDl46O1bG0D3XInv+9tkPrxrN8jUUTiFlDkkmKWgn1M6CfIA13SuGqOa9Korw==",
+			"license": "ISC",
+			"engines": {
+				"node": ">=14"
+			},
+			"funding": {
+				"url": "https://github.com/sponsors/isaacs"
+			}
+		},
 		"node_modules/simple-concat": {
 			"version": "1.0.1",
 			"resolved": "https://registry.npmjs.org/simple-concat/-/simple-concat-1.0.1.tgz",
@@ -4140,7 +4448,21 @@
 			"version": "4.2.3",
 			"resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
 			"integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
-			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"emoji-regex": "^8.0.0",
+				"is-fullwidth-code-point": "^3.0.0",
+				"strip-ansi": "^6.0.1"
+			},
+			"engines": {
+				"node": ">=8"
+			}
+		},
+		"node_modules/string-width-cjs": {
+			"name": "string-width",
+			"version": "4.2.3",
+			"resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
+			"integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
 			"license": "MIT",
 			"dependencies": {
 				"emoji-regex": "^8.0.0",
@@ -4163,6 +4485,19 @@
 				"node": ">=8"
 			}
 		},
+		"node_modules/strip-ansi-cjs": {
+			"name": "strip-ansi",
+			"version": "6.0.1",
+			"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
+			"integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==",
+			"license": "MIT",
+			"dependencies": {
+				"ansi-regex": "^5.0.1"
+			},
+			"engines": {
+				"node": ">=8"
+			}
+		},
 		"node_modules/strip-json-comments": {
 			"version": "2.0.1",
 			"resolved": "https://registry.npmjs.org/strip-json-comments/-/strip-json-comments-2.0.1.tgz",
@@ -4679,6 +5014,21 @@
 				"webidl-conversions": "^3.0.0"
 			}
 		},
+		"node_modules/which": {
+			"version": "2.0.2",
+			"resolved": "https://registry.npmjs.org/which/-/which-2.0.2.tgz",
+			"integrity": "sha512-BLI3Tl1TW3Pvl70l3yq3Y64i+awpwXqsGBYWkkqMtnbXgrMD+yj7rhW0kuEDxzJaYXGjEW5ogapKNMEKNMjibA==",
+			"license": "ISC",
+			"dependencies": {
+				"isexe": "^2.0.0"
+			},
+			"bin": {
+				"node-which": "bin/node-which"
+			},
+			"engines": {
+				"node": ">= 8"
+			}
+		},
 		"node_modules/why-is-node-running": {
 			"version": "2.3.0",
 			"resolved": "https://registry.npmjs.org/why-is-node-running/-/why-is-node-running-2.3.0.tgz",
@@ -4732,6 +5082,24 @@
 				"url": "https://github.com/chalk/wrap-ansi?sponsor=1"
 			}
 		},
+		"node_modules/wrap-ansi-cjs": {
+			"name": "wrap-ansi",
+			"version": "7.0.0",
+			"resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-7.0.0.tgz",
+			"integrity": "sha512-YVGIj2kamLSTxw6NsZjoBxfSwsn0ycdesmc4p+Q21c5zPuZ1pl+NfxVdxPtdHvmNVOQ6XSYG4AUtyt/Fi7D16Q==",
+			"license": "MIT",
+			"dependencies": {
+				"ansi-styles": "^4.0.0",
+				"string-width": "^4.1.0",
+				"strip-ansi": "^6.0.0"
+			},
+			"engines": {
+				"node": ">=10"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/wrap-ansi?sponsor=1"
+			}
+		},
 		"node_modules/wrappy": {
 			"version": "1.0.2",
 			"resolved": "https://registry.npmjs.org/wrappy/-/wrappy-1.0.2.tgz",
@@ -4853,6 +5221,22 @@
 			"name": "@mariozechner/pi-agent",
 			"version": "0.5.44",
 			"license": "MIT",
+			"dependencies": {
+				"@mariozechner/pi-ai": "^0.5.44"
+			},
+			"devDependencies": {
+				"@types/node": "^24.3.0",
+				"typescript": "^5.7.3",
+				"vitest": "^3.2.4"
+			},
+			"engines": {
+				"node": ">=20.0.0"
+			}
+		},
+		"packages/agent-old": {
+			"name": "@mariozechner/pi-agent-old",
+			"version": "0.5.44",
+			"license": "MIT",
 			"dependencies": {
 				"@mariozechner/pi-tui": "^0.5.44",
 				"@types/glob": "^8.1.0",
@@ -4868,78 +5252,7 @@
 				"node": ">=20.0.0"
 			}
 		},
-		"packages/agent/node_modules/@isaacs/balanced-match": {
-			"version": "4.0.1",
-			"license": "MIT",
-			"engines": {
-				"node": "20 || >=22"
-			}
-		},
-		"packages/agent/node_modules/@isaacs/brace-expansion": {
-			"version": "5.0.0",
-			"license": "MIT",
-			"dependencies": {
-				"@isaacs/balanced-match": "^4.0.1"
-			},
-			"engines": {
-				"node": "20 || >=22"
-			}
-		},
-		"packages/agent/node_modules/@isaacs/cliui": {
-			"version": "8.0.2",
-			"license": "ISC",
-			"dependencies": {
-				"string-width": "^5.1.2",
-				"string-width-cjs": "npm:string-width@^4.2.0",
-				"strip-ansi": "^7.0.1",
-				"strip-ansi-cjs": "npm:strip-ansi@^6.0.1",
-				"wrap-ansi": "^8.1.0",
-				"wrap-ansi-cjs": "npm:wrap-ansi@^7.0.0"
-			},
-			"engines": {
-				"node": ">=12"
-			}
-		},
-		"packages/agent/node_modules/@isaacs/cliui/node_modules/ansi-regex": {
-			"version": "6.2.2",
-			"resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.2.2.tgz",
-			"integrity": "sha512-Bq3SmSpyFHaWjPk8If9yc6svM8c56dB5BAtW4Qbw5jHTwwXXcTLoRMkpDJp6VL0XzlWaCHTXrkFURMYmD0sLqg==",
-			"license": "MIT",
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/ansi-regex?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/@isaacs/cliui/node_modules/strip-ansi": {
-			"version": "7.1.2",
-			"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.1.2.tgz",
-			"integrity": "sha512-gmBGslpoQJtgnMAvOVqGZpEz9dyoKTCzy2nfz/n8aIFhN/jCE/rCmcxabB6jOOHV+0WNnylOxaxBQPSvcWklhA==",
-			"license": "MIT",
-			"dependencies": {
-				"ansi-regex": "^6.0.1"
-			},
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/strip-ansi?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/@types/glob": {
-			"version": "8.1.0",
-			"license": "MIT",
-			"dependencies": {
-				"@types/minimatch": "^5.1.2",
-				"@types/node": "*"
-			}
-		},
-		"packages/agent/node_modules/@types/minimatch": {
-			"version": "5.1.2",
-			"license": "MIT"
-		},
-		"packages/agent/node_modules/chalk": {
+		"packages/agent-old/node_modules/chalk": {
 			"version": "5.6.2",
 			"resolved": "https://registry.npmjs.org/chalk/-/chalk-5.6.2.tgz",
 			"integrity": "sha512-7NzBL0rN6fMUW+f7A6Io4h40qQlG+xGmtMxfbnH/K7TAtt8JQWVQK+6g0UXKMeVJoyV5EkkNsErQ8pVD3bLHbA==",
@@ -4951,309 +5264,23 @@
 				"url": "https://github.com/chalk/chalk?sponsor=1"
 			}
 		},
-		"packages/agent/node_modules/cross-spawn": {
-			"version": "7.0.6",
+		"packages/agent/node_modules/@types/node": {
+			"version": "24.8.0",
+			"resolved": "https://registry.npmjs.org/@types/node/-/node-24.8.0.tgz",
+			"integrity": "sha512-5x08bUtU8hfboMTrJ7mEO4CpepS9yBwAqcL52y86SWNmbPX8LVbNs3EP4cNrIZgdjk2NAlP2ahNihozpoZIxSg==",
+			"dev": true,
 			"license": "MIT",
 			"dependencies": {
-				"path-key": "^3.1.0",
-				"shebang-command": "^2.0.0",
-				"which": "^2.0.1"
-			},
-			"engines": {
-				"node": ">= 8"
+				"undici-types": "~7.14.0"
 			}
 		},
-		"packages/agent/node_modules/eastasianwidth": {
-			"version": "0.2.0",
+		"packages/agent/node_modules/undici-types": {
+			"version": "7.14.0",
+			"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.14.0.tgz",
+			"integrity": "sha512-QQiYxHuyZ9gQUIrmPo3IA+hUl4KYk8uSA7cHrcKd/l3p1OTpZcM0Tbp9x7FAtXdAYhlasd60ncPpgu6ihG6TOA==",
+			"dev": true,
 			"license": "MIT"
 		},
-		"packages/agent/node_modules/emoji-regex": {
-			"version": "9.2.2",
-			"license": "MIT"
-		},
-		"packages/agent/node_modules/foreground-child": {
-			"version": "3.3.1",
-			"license": "ISC",
-			"dependencies": {
-				"cross-spawn": "^7.0.6",
-				"signal-exit": "^4.0.1"
-			},
-			"engines": {
-				"node": ">=14"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/isaacs"
-			}
-		},
-		"packages/agent/node_modules/glob": {
-			"version": "11.0.3",
-			"license": "ISC",
-			"dependencies": {
-				"foreground-child": "^3.3.1",
-				"jackspeak": "^4.1.1",
-				"minimatch": "^10.0.3",
-				"minipass": "^7.1.2",
-				"package-json-from-dist": "^1.0.0",
-				"path-scurry": "^2.0.0"
-			},
-			"bin": {
-				"glob": "dist/esm/bin.mjs"
-			},
-			"engines": {
-				"node": "20 || >=22"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/isaacs"
-			}
-		},
-		"packages/agent/node_modules/isexe": {
-			"version": "2.0.0",
-			"license": "ISC"
-		},
-		"packages/agent/node_modules/jackspeak": {
-			"version": "4.1.1",
-			"license": "BlueOak-1.0.0",
-			"dependencies": {
-				"@isaacs/cliui": "^8.0.2"
-			},
-			"engines": {
-				"node": "20 || >=22"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/isaacs"
-			}
-		},
-		"packages/agent/node_modules/lru-cache": {
-			"version": "11.1.0",
-			"license": "ISC",
-			"engines": {
-				"node": "20 || >=22"
-			}
-		},
-		"packages/agent/node_modules/minimatch": {
-			"version": "10.0.3",
-			"license": "ISC",
-			"dependencies": {
-				"@isaacs/brace-expansion": "^5.0.0"
-			},
-			"engines": {
-				"node": "20 || >=22"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/isaacs"
-			}
-		},
-		"packages/agent/node_modules/package-json-from-dist": {
-			"version": "1.0.1",
-			"license": "BlueOak-1.0.0"
-		},
-		"packages/agent/node_modules/path-key": {
-			"version": "3.1.1",
-			"license": "MIT",
-			"engines": {
-				"node": ">=8"
-			}
-		},
-		"packages/agent/node_modules/path-scurry": {
-			"version": "2.0.0",
-			"license": "BlueOak-1.0.0",
-			"dependencies": {
-				"lru-cache": "^11.0.0",
-				"minipass": "^7.1.2"
-			},
-			"engines": {
-				"node": "20 || >=22"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/isaacs"
-			}
-		},
-		"packages/agent/node_modules/shebang-command": {
-			"version": "2.0.0",
-			"license": "MIT",
-			"dependencies": {
-				"shebang-regex": "^3.0.0"
-			},
-			"engines": {
-				"node": ">=8"
-			}
-		},
-		"packages/agent/node_modules/shebang-regex": {
-			"version": "3.0.0",
-			"license": "MIT",
-			"engines": {
-				"node": ">=8"
-			}
-		},
-		"packages/agent/node_modules/string-width": {
-			"version": "5.1.2",
-			"license": "MIT",
-			"dependencies": {
-				"eastasianwidth": "^0.2.0",
-				"emoji-regex": "^9.2.2",
-				"strip-ansi": "^7.0.1"
-			},
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/sponsors/sindresorhus"
-			}
-		},
-		"packages/agent/node_modules/string-width-cjs": {
-			"name": "string-width",
-			"version": "4.2.3",
-			"license": "MIT",
-			"dependencies": {
-				"emoji-regex": "^8.0.0",
-				"is-fullwidth-code-point": "^3.0.0",
-				"strip-ansi": "^6.0.1"
-			},
-			"engines": {
-				"node": ">=8"
-			}
-		},
-		"packages/agent/node_modules/string-width-cjs/node_modules/emoji-regex": {
-			"version": "8.0.0",
-			"license": "MIT"
-		},
-		"packages/agent/node_modules/string-width/node_modules/ansi-regex": {
-			"version": "6.2.2",
-			"resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.2.2.tgz",
-			"integrity": "sha512-Bq3SmSpyFHaWjPk8If9yc6svM8c56dB5BAtW4Qbw5jHTwwXXcTLoRMkpDJp6VL0XzlWaCHTXrkFURMYmD0sLqg==",
-			"license": "MIT",
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/ansi-regex?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/string-width/node_modules/strip-ansi": {
-			"version": "7.1.2",
-			"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.1.2.tgz",
-			"integrity": "sha512-gmBGslpoQJtgnMAvOVqGZpEz9dyoKTCzy2nfz/n8aIFhN/jCE/rCmcxabB6jOOHV+0WNnylOxaxBQPSvcWklhA==",
-			"license": "MIT",
-			"dependencies": {
-				"ansi-regex": "^6.0.1"
-			},
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/strip-ansi?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/strip-ansi-cjs": {
-			"name": "strip-ansi",
-			"version": "6.0.1",
-			"license": "MIT",
-			"dependencies": {
-				"ansi-regex": "^5.0.1"
-			},
-			"engines": {
-				"node": ">=8"
-			}
-		},
-		"packages/agent/node_modules/which": {
-			"version": "2.0.2",
-			"license": "ISC",
-			"dependencies": {
-				"isexe": "^2.0.0"
-			},
-			"bin": {
-				"node-which": "bin/node-which"
-			},
-			"engines": {
-				"node": ">= 8"
-			}
-		},
-		"packages/agent/node_modules/wrap-ansi": {
-			"version": "8.1.0",
-			"license": "MIT",
-			"dependencies": {
-				"ansi-styles": "^6.1.0",
-				"string-width": "^5.0.1",
-				"strip-ansi": "^7.0.1"
-			},
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/wrap-ansi?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/wrap-ansi-cjs": {
-			"name": "wrap-ansi",
-			"version": "7.0.0",
-			"license": "MIT",
-			"dependencies": {
-				"ansi-styles": "^4.0.0",
-				"string-width": "^4.1.0",
-				"strip-ansi": "^6.0.0"
-			},
-			"engines": {
-				"node": ">=10"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/wrap-ansi?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/wrap-ansi-cjs/node_modules/emoji-regex": {
-			"version": "8.0.0",
-			"license": "MIT"
-		},
-		"packages/agent/node_modules/wrap-ansi-cjs/node_modules/string-width": {
-			"version": "4.2.3",
-			"license": "MIT",
-			"dependencies": {
-				"emoji-regex": "^8.0.0",
-				"is-fullwidth-code-point": "^3.0.0",
-				"strip-ansi": "^6.0.1"
-			},
-			"engines": {
-				"node": ">=8"
-			}
-		},
-		"packages/agent/node_modules/wrap-ansi/node_modules/ansi-regex": {
-			"version": "6.2.2",
-			"resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.2.2.tgz",
-			"integrity": "sha512-Bq3SmSpyFHaWjPk8If9yc6svM8c56dB5BAtW4Qbw5jHTwwXXcTLoRMkpDJp6VL0XzlWaCHTXrkFURMYmD0sLqg==",
-			"license": "MIT",
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/ansi-regex?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/wrap-ansi/node_modules/ansi-styles": {
-			"version": "6.2.3",
-			"resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-6.2.3.tgz",
-			"integrity": "sha512-4Dj6M28JB+oAH8kFkTLUo+a2jwOFkuqb3yucU0CANcRRUbxS0cP0nZYCGjcc3BNXwRIsUVmDGgzawme7zvJHvg==",
-			"license": "MIT",
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/ansi-styles?sponsor=1"
-			}
-		},
-		"packages/agent/node_modules/wrap-ansi/node_modules/strip-ansi": {
-			"version": "7.1.2",
-			"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.1.2.tgz",
-			"integrity": "sha512-gmBGslpoQJtgnMAvOVqGZpEz9dyoKTCzy2nfz/n8aIFhN/jCE/rCmcxabB6jOOHV+0WNnylOxaxBQPSvcWklhA==",
-			"license": "MIT",
-			"dependencies": {
-				"ansi-regex": "^6.0.1"
-			},
-			"engines": {
-				"node": ">=12"
-			},
-			"funding": {
-				"url": "https://github.com/chalk/strip-ansi?sponsor=1"
-			}
-		},
 		"packages/ai": {
 			"name": "@mariozechner/pi-ai",
 			"version": "0.5.44",
@@ -5351,12 +5378,64 @@
 				"ws": "^8.18.0"
 			}
 		},
+		"packages/coding-agent": {
+			"name": "@mariozechner/coding-agent",
+			"version": "0.5.44",
+			"license": "MIT",
+			"dependencies": {
+				"@mariozechner/pi-agent": "^0.5.44",
+				"@mariozechner/pi-ai": "^0.5.44",
+				"@mariozechner/pi-tui": "^0.5.44",
+				"chalk": "^5.5.0",
+				"glob": "^11.0.3"
+			},
+			"bin": {
+				"coding-agent": "dist/cli.js"
+			},
+			"devDependencies": {
+				"@types/node": "^24.3.0",
+				"typescript": "^5.7.3",
+				"vitest": "^3.2.4"
+			},
+			"engines": {
+				"node": ">=20.0.0"
+			}
+		},
+		"packages/coding-agent/node_modules/@types/node": {
+			"version": "24.8.0",
+			"resolved": "https://registry.npmjs.org/@types/node/-/node-24.8.0.tgz",
+			"integrity": "sha512-5x08bUtU8hfboMTrJ7mEO4CpepS9yBwAqcL52y86SWNmbPX8LVbNs3EP4cNrIZgdjk2NAlP2ahNihozpoZIxSg==",
+			"dev": true,
+			"license": "MIT",
+			"dependencies": {
+				"undici-types": "~7.14.0"
+			}
+		},
+		"packages/coding-agent/node_modules/chalk": {
+			"version": "5.6.2",
+			"resolved": "https://registry.npmjs.org/chalk/-/chalk-5.6.2.tgz",
+			"integrity": "sha512-7NzBL0rN6fMUW+f7A6Io4h40qQlG+xGmtMxfbnH/K7TAtt8JQWVQK+6g0UXKMeVJoyV5EkkNsErQ8pVD3bLHbA==",
+			"license": "MIT",
+			"engines": {
+				"node": "^12.17.0 || ^14.13 || >=16.0.0"
+			},
+			"funding": {
+				"url": "https://github.com/chalk/chalk?sponsor=1"
+			}
+		},
+		"packages/coding-agent/node_modules/undici-types": {
+			"version": "7.14.0",
+			"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.14.0.tgz",
+			"integrity": "sha512-QQiYxHuyZ9gQUIrmPo3IA+hUl4KYk8uSA7cHrcKd/l3p1OTpZcM0Tbp9x7FAtXdAYhlasd60ncPpgu6ihG6TOA==",
+			"dev": true,
+			"license": "MIT"
+		},
 		"packages/pods": {
 			"name": "@mariozechner/pi",
 			"version": "0.5.44",
 			"license": "MIT",
 			"dependencies": {
-				"@mariozechner/pi-agent": "^0.5.44",
+				"@mariozechner/pi-agent-old": "^0.5.44",
 				"chalk": "^5.5.0"
 			},
 			"bin": {

package.json

@@ -7,7 +7,7 @@
 	],
 	"scripts": {
 		"clean": "npm run clean --workspaces",
-		"build": "npm run build -w @mariozechner/pi-tui && npm run build -w @mariozechner/pi-ai && npm run build -w @mariozechner/pi-web-ui && npm run build -w @mariozechner/pi-agent && npm run build -w @mariozechner/pi-proxy && npm run build -w @mariozechner/pi",
+		"build": "npm run build -w @mariozechner/pi-tui && npm run build -w @mariozechner/pi-ai && npm run build -w @mariozechner/pi-agent && npm run build -w @mariozechner/pi-agent-old && npm run build -w @mariozechner/coding-agent && npm run build -w @mariozechner/pi-web-ui && npm run build -w @mariozechner/pi-proxy && npm run build -w @mariozechner/pi",
 		"dev": "concurrently --names \"ai,web-ui,tui,proxy\" --prefix-colors \"cyan,green,magenta,blue\" \"npm run dev -w @mariozechner/pi-ai\" \"npm run dev -w @mariozechner/pi-web-ui\" \"npm run dev -w @mariozechner/pi-tui\" \"npm run dev -w @mariozechner/pi-proxy\"",
 		"check": "biome check --write . && npm run check --workspaces && tsc --noEmit",
 		"test": "npm run test --workspaces --if-present",

packages/agent-old/package.json

@@ -0,0 +1,47 @@
+{
+	"name": "@mariozechner/pi-agent-old",
+	"version": "0.5.44",
+	"description": "General-purpose agent with tool calling and session persistence",
+	"type": "module",
+	"bin": {
+		"pi-agent": "dist/cli.js"
+	},
+	"main": "./dist/index.js",
+	"types": "./dist/index.d.ts",
+	"files": [
+		"dist"
+	],
+	"scripts": {
+		"clean": "rm -rf dist",
+		"build": "tsc -p tsconfig.build.json && chmod +x dist/cli.js",
+		"check": "biome check --write .",
+		"prepublishOnly": "npm run clean && npm run build"
+	},
+	"dependencies": {
+		"@mariozechner/pi-tui": "^0.5.44",
+		"@types/glob": "^8.1.0",
+		"chalk": "^5.5.0",
+		"glob": "^11.0.3",
+		"openai": "^5.12.2"
+	},
+	"devDependencies": {},
+	"keywords": [
+		"agent",
+		"ai",
+		"llm",
+		"openai",
+		"claude",
+		"cli",
+		"tui"
+	],
+	"author": "Mario Zechner",
+	"license": "MIT",
+	"repository": {
+		"type": "git",
+		"url": "git+https://github.com/badlogic/pi-mono.git",
+		"directory": "packages/agent"
+	},
+	"engines": {
+		"node": ">=20.0.0"
+	}
+}

packages/agent-old/src/agent.ts

@@ -0,0 +1,741 @@
+import OpenAI from "openai";
+import type { ResponseFunctionToolCallOutputItem } from "openai/resources/responses/responses.mjs";
+import type { SessionManager } from "./session-manager.js";
+import { executeTool, toolsForChat, toolsForResponses } from "./tools/tools.js";
+
+export type AgentEvent =
+	| { type: "session_start"; sessionId: string; model: string; api: string; baseURL: string; systemPrompt: string }
+	| { type: "assistant_start" }
+	| { type: "reasoning"; text: string }
+	| { type: "tool_call"; toolCallId: string; name: string; args: string }
+	| { type: "tool_result"; toolCallId: string; result: string; isError: boolean }
+	| { type: "assistant_message"; text: string }
+	| { type: "error"; message: string }
+	| { type: "user_message"; text: string }
+	| { type: "interrupted" }
+	| {
+			type: "token_usage";
+			inputTokens: number;
+			outputTokens: number;
+			totalTokens: number;
+			cacheReadTokens: number;
+			cacheWriteTokens: number;
+			reasoningTokens: number;
+	  };
+
+export interface AgentEventReceiver {
+	on(event: AgentEvent): Promise<void>;
+}
+
+export interface AgentConfig {
+	apiKey: string;
+	baseURL: string;
+	model: string;
+	api: "completions" | "responses";
+	systemPrompt: string;
+}
+
+export interface ToolCall {
+	name: string;
+	arguments: string;
+	id: string;
+}
+
+// Cache for model reasoning support detection per API type
+const modelReasoningSupport = new Map<string, { completions?: boolean; responses?: boolean }>();
+
+// Provider detection based on base URL
+function detectProvider(baseURL?: string): "openai" | "gemini" | "groq" | "anthropic" | "openrouter" | "other" {
+	if (!baseURL) return "openai";
+	if (baseURL.includes("api.openai.com")) return "openai";
+	if (baseURL.includes("generativelanguage.googleapis.com")) return "gemini";
+	if (baseURL.includes("api.groq.com")) return "groq";
+	if (baseURL.includes("api.anthropic.com")) return "anthropic";
+	if (baseURL.includes("openrouter.ai")) return "openrouter";
+	return "other";
+}
+
+// Parse provider-specific reasoning from message content
+function parseReasoningFromMessage(message: any, baseURL?: string): { cleanContent: string; reasoningTexts: string[] } {
+	const provider = detectProvider(baseURL);
+	const reasoningTexts: string[] = [];
+	let cleanContent = message.content || "";
+
+	switch (provider) {
+		case "gemini":
+			// Gemini returns thinking in <thought> tags
+			if (cleanContent.includes("<thought>")) {
+				const thoughtMatches = cleanContent.matchAll(/<thought>([\s\S]*?)<\/thought>/g);
+				for (const match of thoughtMatches) {
+					reasoningTexts.push(match[1].trim());
+				}
+				// Remove all thought tags from the response
+				cleanContent = cleanContent.replace(/<thought>[\s\S]*?<\/thought>/g, "").trim();
+			}
+			break;
+
+		case "groq":
+			// Groq returns reasoning in a separate field when reasoning_format is "parsed"
+			if (message.reasoning) {
+				reasoningTexts.push(message.reasoning);
+			}
+			break;
+
+		case "openrouter":
+			// OpenRouter returns reasoning in message.reasoning field
+			if (message.reasoning) {
+				reasoningTexts.push(message.reasoning);
+			}
+			break;
+
+		default:
+			// Other providers don't embed reasoning in message content
+			break;
+	}
+
+	return { cleanContent, reasoningTexts };
+}
+
+// Adjust request options based on provider-specific requirements
+function adjustRequestForProvider(
+	requestOptions: any,
+	api: "completions" | "responses",
+	baseURL?: string,
+	supportsReasoning?: boolean,
+): any {
+	const provider = detectProvider(baseURL);
+
+	// Handle provider-specific adjustments
+	switch (provider) {
+		case "gemini":
+			if (api === "completions" && supportsReasoning && requestOptions.reasoning_effort) {
+				// Gemini needs extra_body for thinking content
+				// Can't use both reasoning_effort and thinking_config
+				const budget =
+					requestOptions.reasoning_effort === "low"
+						? 1024
+						: requestOptions.reasoning_effort === "medium"
+							? 8192
+							: 24576;
+
+				requestOptions.extra_body = {
+					google: {
+						thinking_config: {
+							thinking_budget: budget,
+							include_thoughts: true,
+						},
+					},
+				};
+				// Remove reasoning_effort when using thinking_config
+				delete requestOptions.reasoning_effort;
+			}
+			break;
+
+		case "groq":
+			if (api === "responses" && requestOptions.reasoning) {
+				// Groq responses API doesn't support reasoning.summary
+				delete requestOptions.reasoning.summary;
+			} else if (api === "completions" && supportsReasoning && requestOptions.reasoning_effort) {
+				// Groq Chat Completions uses reasoning_format instead of reasoning_effort alone
+				requestOptions.reasoning_format = "parsed";
+				// Keep reasoning_effort for Groq
+			}
+			break;
+
+		case "anthropic":
+			// Anthropic's OpenAI compatibility has its own quirks
+			// But thinking content isn't available via OpenAI compat layer
+			break;
+
+		case "openrouter":
+			// OpenRouter uses a unified reasoning parameter format
+			if (api === "completions" && supportsReasoning && requestOptions.reasoning_effort) {
+				// Convert reasoning_effort to OpenRouter's reasoning format
+				requestOptions.reasoning = {
+					effort:
+						requestOptions.reasoning_effort === "low"
+							? "low"
+							: requestOptions.reasoning_effort === "minimal"
+								? "low"
+								: requestOptions.reasoning_effort === "medium"
+									? "medium"
+									: "high",
+				};
+				delete requestOptions.reasoning_effort;
+			}
+			break;
+
+		default:
+			// OpenAI and others use standard format
+			break;
+	}
+
+	return requestOptions;
+}
+
+async function checkReasoningSupport(
+	client: OpenAI,
+	model: string,
+	api: "completions" | "responses",
+	baseURL?: string,
+	signal?: AbortSignal,
+): Promise<boolean> {
+	// Check if already aborted
+	if (signal?.aborted) {
+		throw new Error("Interrupted");
+	}
+
+	// Check cache first
+	const cacheKey = model;
+	const cached = modelReasoningSupport.get(cacheKey);
+	if (cached && cached[api] !== undefined) {
+		return cached[api]!;
+	}
+
+	let supportsReasoning = false;
+	const provider = detectProvider(baseURL);
+
+	if (api === "responses") {
+		// Try a minimal request with reasoning parameter for Responses API
+		try {
+			const testRequest: any = {
+				model,
+				input: "test",
+				max_output_tokens: 1024,
+				reasoning: {
+					effort: "low", // Use low instead of minimal to ensure we get summaries
+				},
+			};
+			await client.responses.create(testRequest, { signal });
+			supportsReasoning = true;
+		} catch (error) {
+			supportsReasoning = false;
+		}
+	} else {
+		// For Chat Completions API, try with reasoning parameter
+		try {
+			const testRequest: any = {
+				model,
+				messages: [{ role: "user", content: "test" }],
+				max_completion_tokens: 1024,
+			};
+
+			// Add provider-specific reasoning parameters
+			if (provider === "gemini") {
+				// Gemini uses extra_body for thinking
+				testRequest.extra_body = {
+					google: {
+						thinking_config: {
+							thinking_budget: 100, // Minimum viable budget for test
+							include_thoughts: true,
+						},
+					},
+				};
+			} else if (provider === "groq") {
+				// Groq uses both reasoning_format and reasoning_effort
+				testRequest.reasoning_format = "parsed";
+				testRequest.reasoning_effort = "low";
+			} else {
+				// Others use reasoning_effort
+				testRequest.reasoning_effort = "minimal";
+			}
+
+			await client.chat.completions.create(testRequest, { signal });
+			supportsReasoning = true;
+		} catch (error) {
+			supportsReasoning = false;
+		}
+	}
+
+	// Update cache
+	const existing = modelReasoningSupport.get(cacheKey) || {};
+	existing[api] = supportsReasoning;
+	modelReasoningSupport.set(cacheKey, existing);
+
+	return supportsReasoning;
+}
+
+export async function callModelResponsesApi(
+	client: OpenAI,
+	model: string,
+	messages: any[],
+	signal?: AbortSignal,
+	eventReceiver?: AgentEventReceiver,
+	supportsReasoning?: boolean,
+	baseURL?: string,
+): Promise<void> {
+	let conversationDone = false;
+
+	while (!conversationDone) {
+		// Check if we've been interrupted
+		if (signal?.aborted) {
+			throw new Error("Interrupted");
+		}
+
+		// Build request options
+		let requestOptions: any = {
+			model,
+			input: messages,
+			tools: toolsForResponses as any,
+			tool_choice: "auto",
+			parallel_tool_calls: true,
+			max_output_tokens: 2000, // TODO make configurable
+			...(supportsReasoning && {
+				reasoning: {
+					effort: "minimal", // Use minimal effort for responses API
+					summary: "detailed", // Request detailed reasoning summaries
+				},
+			}),
+		};
+
+		// Apply provider-specific adjustments
+		requestOptions = adjustRequestForProvider(requestOptions, "responses", baseURL, supportsReasoning);
+
+		const response = await client.responses.create(requestOptions, { signal });
+
+		// Report token usage if available (responses API format)
+		if (response.usage) {
+			const usage = response.usage;
+			eventReceiver?.on({
+				type: "token_usage",
+				inputTokens: usage.input_tokens || 0,
+				outputTokens: usage.output_tokens || 0,
+				totalTokens: usage.total_tokens || 0,
+				cacheReadTokens: usage.input_tokens_details?.cached_tokens || 0,
+				cacheWriteTokens: 0, // Not available in API
+				reasoningTokens: usage.output_tokens_details?.reasoning_tokens || 0,
+			});
+		}
+
+		const output = response.output;
+		if (!output) break;
+
+		for (const item of output) {
+			// gpt-oss vLLM quirk: need to remove type from "message" events
+			if (item.id === "message") {
+				const { type, ...message } = item;
+				messages.push(item);
+			} else {
+				messages.push(item);
+			}
+
+			switch (item.type) {
+				case "reasoning": {
+					// Handle both content (o1/o3) and summary (gpt-5) formats
+					const reasoningItems = item.content || item.summary || [];
+					for (const content of reasoningItems) {
+						if (content.type === "reasoning_text" || content.type === "summary_text") {
+							await eventReceiver?.on({ type: "reasoning", text: content.text });
+						}
+					}
+					break;
+				}
+
+				case "message": {
+					for (const content of item.content || []) {
+						if (content.type === "output_text") {
+							await eventReceiver?.on({ type: "assistant_message", text: content.text });
+						} else if (content.type === "refusal") {
+							await eventReceiver?.on({ type: "error", message: `Refusal: ${content.refusal}` });
+						}
+						conversationDone = true;
+					}
+					break;
+				}
+
+				case "function_call": {
+					if (signal?.aborted) {
+						throw new Error("Interrupted");
+					}
+
+					try {
+						await eventReceiver?.on({
+							type: "tool_call",
+							toolCallId: item.call_id || "",
+							name: item.name,
+							args: item.arguments,
+						});
+						const result = await executeTool(item.name, item.arguments, signal);
+						await eventReceiver?.on({
+							type: "tool_result",
+							toolCallId: item.call_id || "",
+							result,
+							isError: false,
+						});
+
+						// Add tool result to messages
+						const toolResultMsg = {
+							type: "function_call_output",
+							call_id: item.call_id,
+							output: result,
+						} as ResponseFunctionToolCallOutputItem;
+						messages.push(toolResultMsg);
+					} catch (e: any) {
+						await eventReceiver?.on({
+							type: "tool_result",
+							toolCallId: item.call_id || "",
+							result: e.message,
+							isError: true,
+						});
+						const errorMsg = {
+							type: "function_call_output",
+							call_id: item.id,
+							output: e.message,
+							isError: true,
+						};
+						messages.push(errorMsg);
+					}
+					break;
+				}
+
+				default: {
+					eventReceiver?.on({ type: "error", message: `Unknown output type in LLM response: ${item.type}` });
+					break;
+				}
+			}
+		}
+	}
+}
+
+export async function callModelChatCompletionsApi(
+	client: OpenAI,
+	model: string,
+	messages: any[],
+	signal?: AbortSignal,
+	eventReceiver?: AgentEventReceiver,
+	supportsReasoning?: boolean,
+	baseURL?: string,
+): Promise<void> {
+	let assistantResponded = false;
+
+	while (!assistantResponded) {
+		if (signal?.aborted) {
+			throw new Error("Interrupted");
+		}
+
+		// Build request options
+		let requestOptions: any = {
+			model,
+			messages,
+			tools: toolsForChat,
+			tool_choice: "auto",
+			max_completion_tokens: 2000, // TODO make configurable
+			...(supportsReasoning && {
+				reasoning_effort: "low", // Use low effort for completions API
+			}),
+		};
+
+		// Apply provider-specific adjustments
+		requestOptions = adjustRequestForProvider(requestOptions, "completions", baseURL, supportsReasoning);
+
+		const response = await client.chat.completions.create(requestOptions, { signal });
+
+		const message = response.choices[0].message;
+
+		// Report token usage if available
+		if (response.usage) {
+			const usage = response.usage;
+			await eventReceiver?.on({
+				type: "token_usage",
+				inputTokens: usage.prompt_tokens || 0,
+				outputTokens: usage.completion_tokens || 0,
+				totalTokens: usage.total_tokens || 0,
+				cacheReadTokens: usage.prompt_tokens_details?.cached_tokens || 0,
+				cacheWriteTokens: 0, // Not available in API
+				reasoningTokens: usage.completion_tokens_details?.reasoning_tokens || 0,
+			});
+		}
+
+		if (message.tool_calls && message.tool_calls.length > 0) {
+			// Add assistant message with tool calls to history
+			const assistantMsg: any = {
+				role: "assistant",
+				content: message.content || null,
+				tool_calls: message.tool_calls,
+			};
+			messages.push(assistantMsg);
+
+			// Display and execute each tool call
+			for (const toolCall of message.tool_calls) {
+				// Check if interrupted before executing tool
+				if (signal?.aborted) {
+					throw new Error("Interrupted");
+				}
+
+				try {
+					const funcName = toolCall.type === "function" ? toolCall.function.name : toolCall.custom.name;
+					const funcArgs = toolCall.type === "function" ? toolCall.function.arguments : toolCall.custom.input;
+
+					await eventReceiver?.on({ type: "tool_call", toolCallId: toolCall.id, name: funcName, args: funcArgs });
+					const result = await executeTool(funcName, funcArgs, signal);
+					await eventReceiver?.on({ type: "tool_result", toolCallId: toolCall.id, result, isError: false });
+
+					// Add tool result to messages
+					const toolMsg = {
+						role: "tool",
+						tool_call_id: toolCall.id,
+						content: result,
+					};
+					messages.push(toolMsg);
+				} catch (e: any) {
+					eventReceiver?.on({ type: "tool_result", toolCallId: toolCall.id, result: e.message, isError: true });
+					const errorMsg = {
+						role: "tool",
+						tool_call_id: toolCall.id,
+						content: e.message,
+					};
+					messages.push(errorMsg);
+				}
+			}
+		} else if (message.content) {
+			// Parse provider-specific reasoning from message
+			const { cleanContent, reasoningTexts } = parseReasoningFromMessage(message, baseURL);
+
+			// Emit reasoning events if any
+			for (const reasoning of reasoningTexts) {
+				await eventReceiver?.on({ type: "reasoning", text: reasoning });
+			}
+
+			// Emit the cleaned assistant message
+			await eventReceiver?.on({ type: "assistant_message", text: cleanContent });
+			const finalMsg = { role: "assistant", content: cleanContent };
+			messages.push(finalMsg);
+			assistantResponded = true;
+		}
+	}
+}
+
+export class Agent {
+	private client: OpenAI;
+	public readonly config: AgentConfig;
+	private messages: any[] = [];
+	private renderer?: AgentEventReceiver;
+	private sessionManager?: SessionManager;
+	private comboReceiver: AgentEventReceiver;
+	private abortController: AbortController | null = null;
+	private supportsReasoning: boolean | null = null;
+
+	constructor(config: AgentConfig, renderer?: AgentEventReceiver, sessionManager?: SessionManager) {
+		this.config = config;
+		this.client = new OpenAI({
+			apiKey: config.apiKey,
+			baseURL: config.baseURL,
+		});
+
+		// Use provided renderer or default to console
+		this.renderer = renderer;
+		this.sessionManager = sessionManager;
+
+		this.comboReceiver = {
+			on: async (event: AgentEvent): Promise<void> => {
+				await this.renderer?.on(event);
+				await this.sessionManager?.on(event);
+			},
+		};
+
+		// Initialize with system prompt if provided
+		if (config.systemPrompt) {
+			this.messages.push({
+				role: "developer",
+				content: config.systemPrompt,
+			});
+		}
+
+		// Start session logging if we have a session manager
+		if (sessionManager) {
+			sessionManager.startSession(this.config);
+
+			// Emit session_start event
+			this.comboReceiver.on({
+				type: "session_start",
+				sessionId: sessionManager.getSessionId(),
+				model: config.model,
+				api: config.api,
+				baseURL: config.baseURL,
+				systemPrompt: config.systemPrompt,
+			});
+		}
+	}
+
+	async ask(userMessage: string): Promise<void> {
+		// Render user message through the event system
+		this.comboReceiver.on({ type: "user_message", text: userMessage });
+
+		// Add user message
+		const userMsg = { role: "user", content: userMessage };
+		this.messages.push(userMsg);
+
+		// Create a new AbortController for this chat session
+		this.abortController = new AbortController();
+
+		try {
+			await this.comboReceiver.on({ type: "assistant_start" });
+
+			// Check reasoning support only once per agent instance
+			if (this.supportsReasoning === null) {
+				this.supportsReasoning = await checkReasoningSupport(
+					this.client,
+					this.config.model,
+					this.config.api,
+					this.config.baseURL,
+					this.abortController.signal,
+				);
+			}
+
+			if (this.config.api === "responses") {
+				await callModelResponsesApi(
+					this.client,
+					this.config.model,
+					this.messages,
+					this.abortController.signal,
+					this.comboReceiver,
+					this.supportsReasoning,
+					this.config.baseURL,
+				);
+			} else {
+				await callModelChatCompletionsApi(
+					this.client,
+					this.config.model,
+					this.messages,
+					this.abortController.signal,
+					this.comboReceiver,
+					this.supportsReasoning,
+					this.config.baseURL,
+				);
+			}
+		} catch (e) {
+			// Check if this was an interruption by checking the abort signal
+			if (this.abortController.signal.aborted) {
+				// Emit interrupted event so UI can clean up properly
+				await this.comboReceiver?.on({ type: "interrupted" });
+				return;
+			}
+			throw e;
+		} finally {
+			this.abortController = null;
+		}
+	}
+
+	interrupt(): void {
+		this.abortController?.abort();
+	}
+
+	setEvents(events: AgentEvent[]): void {
+		// Reconstruct messages from events based on API type
+		this.messages = [];
+
+		if (this.config.api === "responses") {
+			// Responses API format
+			if (this.config.systemPrompt) {
+				this.messages.push({
+					role: "developer",
+					content: this.config.systemPrompt,
+				});
+			}
+
+			for (const event of events) {
+				switch (event.type) {
+					case "user_message":
+						this.messages.push({
+							role: "user",
+							content: [{ type: "input_text", text: event.text }],
+						});
+						break;
+
+					case "reasoning":
+						// Add reasoning message
+						this.messages.push({
+							type: "reasoning",
+							content: [{ type: "reasoning_text", text: event.text }],
+						});
+						break;
+
+					case "tool_call":
+						// Add function call
+						this.messages.push({
+							type: "function_call",
+							id: event.toolCallId,
+							name: event.name,
+							arguments: event.args,
+						});
+						break;
+
+					case "tool_result":
+						// Add function result
+						this.messages.push({
+							type: "function_call_output",
+							call_id: event.toolCallId,
+							output: event.result,
+						});
+						break;
+
+					case "assistant_message":
+						// Add final message
+						this.messages.push({
+							type: "message",
+							content: [{ type: "output_text", text: event.text }],
+						});
+						break;
+				}
+			}
+		} else {
+			// Chat Completions API format
+			if (this.config.systemPrompt) {
+				this.messages.push({ role: "system", content: this.config.systemPrompt });
+			}
+
+			// Track tool calls in progress
+			let pendingToolCalls: any[] = [];
+
+			for (const event of events) {
+				switch (event.type) {
+					case "user_message":
+						this.messages.push({ role: "user", content: event.text });
+						break;
+
+					case "assistant_start":
+						// Reset pending tool calls for new assistant response
+						pendingToolCalls = [];
+						break;
+
+					case "tool_call":
+						// Accumulate tool calls
+						pendingToolCalls.push({
+							id: event.toolCallId,
+							type: "function",
+							function: {
+								name: event.name,
+								arguments: event.args,
+							},
+						});
+						break;
+
+					case "tool_result":
+						// When we see the first tool result, add the assistant message with all tool calls
+						if (pendingToolCalls.length > 0) {
+							this.messages.push({
+								role: "assistant",
+								content: null,
+								tool_calls: pendingToolCalls,
+							});
+							pendingToolCalls = [];
+						}
+						// Add the tool result
+						this.messages.push({
+							role: "tool",
+							tool_call_id: event.toolCallId,
+							content: event.result,
+						});
+						break;
+
+					case "assistant_message":
+						// Final assistant response (no tool calls)
+						this.messages.push({ role: "assistant", content: event.text });
+						break;
+
+					// Skip other event types (thinking, error, interrupted, token_usage)
+				}
+			}
+		}
+	}
+}

packages/agent-old/src/index.ts

@@ -0,0 +1,15 @@
+// Main exports for pi-agent package
+
+export type { AgentConfig, AgentEvent, AgentEventReceiver } from "./agent.js";
+export { Agent } from "./agent.js";
+export type { ArgDef, ArgDefs, ParsedArgs } from "./args.js";
+// CLI utilities
+export { parseArgs, printHelp } from "./args.js";
+// CLI main function
+export { main } from "./main.js";
+// Renderers
+export { ConsoleRenderer } from "./renderers/console-renderer.js";
+export { JsonRenderer } from "./renderers/json-renderer.js";
+export { TuiRenderer } from "./renderers/tui-renderer.js";
+export type { SessionData, SessionEvent, SessionHeader } from "./session-manager.js";
+export { SessionManager } from "./session-manager.js";

packages/agent-old/tsconfig.build.json

@@ -0,0 +1,9 @@
+{
+	"extends": "../../tsconfig.base.json",
+	"compilerOptions": {
+		"outDir": "./dist",
+		"rootDir": "./src"
+	},
+	"include": ["src/**/*"],
+	"exclude": ["node_modules", "dist"]
+}

packages/agent/package.json

@@ -1,38 +1,31 @@
 {
 	"name": "@mariozechner/pi-agent",
 	"version": "0.5.44",
-	"description": "General-purpose agent with tool calling and session persistence",
+	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"type": "module",
-	"bin": {
-		"pi-agent": "dist/cli.js"
-	},
 	"main": "./dist/index.js",
 	"types": "./dist/index.d.ts",
 	"files": [
-		"dist"
+		"dist",
+		"README.md"
 	],
 	"scripts": {
 		"clean": "rm -rf dist",
-		"build": "tsc -p tsconfig.build.json && chmod +x dist/cli.js",
-		"check": "biome check --write .",
+		"build": "tsc -p tsconfig.build.json",
+		"dev": "tsc -p tsconfig.build.json --watch --preserveWatchOutput",
+		"check": "tsc --noEmit",
+		"test": "vitest --run",
 		"prepublishOnly": "npm run clean && npm run build"
 	},
 	"dependencies": {
-		"@mariozechner/pi-tui": "^0.5.44",
-		"@types/glob": "^8.1.0",
-		"chalk": "^5.5.0",
-		"glob": "^11.0.3",
-		"openai": "^5.12.2"
+		"@mariozechner/pi-ai": "^0.5.44"
 	},
-	"devDependencies": {},
 	"keywords": [
-		"agent",
 		"ai",
+		"agent",
 		"llm",
-		"openai",
-		"claude",
-		"cli",
-		"tui"
+		"transport",
+		"state-management"
 	],
 	"author": "Mario Zechner",
 	"license": "MIT",
@@ -43,5 +36,10 @@
 	},
 	"engines": {
 		"node": ">=20.0.0"
+	},
+	"devDependencies": {
+		"@types/node": "^24.3.0",
+		"typescript": "^5.7.3",
+		"vitest": "^3.2.4"
 	}
 }

packages/agent/src/agent.ts

@@ -1,741 +1,283 @@
-import OpenAI from "openai";
-import type { ResponseFunctionToolCallOutputItem } from "openai/resources/responses/responses.mjs";
-import type { SessionManager } from "./session-manager.js";
-import { executeTool, toolsForChat, toolsForResponses } from "./tools/tools.js";
+import type { ImageContent, Message, QueuedMessage, TextContent } from "@mariozechner/pi-ai";
+import { getModel } from "@mariozechner/pi-ai";
+import type { AgentTransport } from "./transports/types.js";
+import type { AgentEvent, AgentState, AppMessage, Attachment, ThinkingLevel } from "./types.js";
 
-export type AgentEvent =
-	| { type: "session_start"; sessionId: string; model: string; api: string; baseURL: string; systemPrompt: string }
-	| { type: "assistant_start" }
-	| { type: "reasoning"; text: string }
-	| { type: "tool_call"; toolCallId: string; name: string; args: string }
-	| { type: "tool_result"; toolCallId: string; result: string; isError: boolean }
-	| { type: "assistant_message"; text: string }
-	| { type: "error"; message: string }
-	| { type: "user_message"; text: string }
-	| { type: "interrupted" }
-	| {
-			type: "token_usage";
-			inputTokens: number;
-			outputTokens: number;
-			totalTokens: number;
-			cacheReadTokens: number;
-			cacheWriteTokens: number;
-			reasoningTokens: number;
-	  };
+/**
+ * Default message transformer: Keep only LLM-compatible messages, strip app-specific fields.
+ * Converts attachments to proper content blocks (images → ImageContent, documents → TextContent).
+ */
+function defaultMessageTransformer(messages: AppMessage[]): Message[] {
+	return messages
+		.filter((m) => {
+			// Only keep standard LLM message roles
+			return m.role === "user" || m.role === "assistant" || m.role === "toolResult";
+		})
+		.map((m) => {
+			if (m.role === "user") {
+				const { attachments, ...rest } = m as any;
 
-export interface AgentEventReceiver {
-	on(event: AgentEvent): Promise<void>;
-}
-
-export interface AgentConfig {
-	apiKey: string;
-	baseURL: string;
-	model: string;
-	api: "completions" | "responses";
-	systemPrompt: string;
-}
-
-export interface ToolCall {
-	name: string;
-	arguments: string;
-	id: string;
-}
-
-// Cache for model reasoning support detection per API type
-const modelReasoningSupport = new Map<string, { completions?: boolean; responses?: boolean }>();
-
-// Provider detection based on base URL
-function detectProvider(baseURL?: string): "openai" | "gemini" | "groq" | "anthropic" | "openrouter" | "other" {
-	if (!baseURL) return "openai";
-	if (baseURL.includes("api.openai.com")) return "openai";
-	if (baseURL.includes("generativelanguage.googleapis.com")) return "gemini";
-	if (baseURL.includes("api.groq.com")) return "groq";
-	if (baseURL.includes("api.anthropic.com")) return "anthropic";
-	if (baseURL.includes("openrouter.ai")) return "openrouter";
-	return "other";
-}
-
-// Parse provider-specific reasoning from message content
-function parseReasoningFromMessage(message: any, baseURL?: string): { cleanContent: string; reasoningTexts: string[] } {
-	const provider = detectProvider(baseURL);
-	const reasoningTexts: string[] = [];
-	let cleanContent = message.content || "";
-
-	switch (provider) {
-		case "gemini":
-			// Gemini returns thinking in <thought> tags
-			if (cleanContent.includes("<thought>")) {
-				const thoughtMatches = cleanContent.matchAll(/<thought>([\s\S]*?)<\/thought>/g);
-				for (const match of thoughtMatches) {
-					reasoningTexts.push(match[1].trim());
-				}
-				// Remove all thought tags from the response
-				cleanContent = cleanContent.replace(/<thought>[\s\S]*?<\/thought>/g, "").trim();
-			}
-			break;
-
-		case "groq":
-			// Groq returns reasoning in a separate field when reasoning_format is "parsed"
-			if (message.reasoning) {
-				reasoningTexts.push(message.reasoning);
-			}
-			break;
-
-		case "openrouter":
-			// OpenRouter returns reasoning in message.reasoning field
-			if (message.reasoning) {
-				reasoningTexts.push(message.reasoning);
-			}
-			break;
-
-		default:
-			// Other providers don't embed reasoning in message content
-			break;
+				// If no attachments, return as-is
+				if (!attachments || attachments.length === 0) {
+					return rest as Message;
 				}
 
-	return { cleanContent, reasoningTexts };
-}
+				// Convert attachments to content blocks
+				const content = Array.isArray(rest.content) ? [...rest.content] : [{ type: "text", text: rest.content }];
 
-// Adjust request options based on provider-specific requirements
-function adjustRequestForProvider(
-	requestOptions: any,
-	api: "completions" | "responses",
-	baseURL?: string,
-	supportsReasoning?: boolean,
-): any {
-	const provider = detectProvider(baseURL);
-
-	// Handle provider-specific adjustments
-	switch (provider) {
-		case "gemini":
-			if (api === "completions" && supportsReasoning && requestOptions.reasoning_effort) {
-				// Gemini needs extra_body for thinking content
-				// Can't use both reasoning_effort and thinking_config
-				const budget =
-					requestOptions.reasoning_effort === "low"
-						? 1024
-						: requestOptions.reasoning_effort === "medium"
-							? 8192
-							: 24576;
-
-				requestOptions.extra_body = {
-					google: {
-						thinking_config: {
-							thinking_budget: budget,
-							include_thoughts: true,
-						},
-					},
-				};
-				// Remove reasoning_effort when using thinking_config
-				delete requestOptions.reasoning_effort;
+				for (const attachment of attachments as Attachment[]) {
+					// Add image blocks for image attachments
+					if (attachment.type === "image") {
+						content.push({
+							type: "image",
+							data: attachment.content,
+							mimeType: attachment.mimeType,
+						} as ImageContent);
 					}
-			break;
-
-		case "groq":
-			if (api === "responses" && requestOptions.reasoning) {
-				// Groq responses API doesn't support reasoning.summary
-				delete requestOptions.reasoning.summary;
-			} else if (api === "completions" && supportsReasoning && requestOptions.reasoning_effort) {
-				// Groq Chat Completions uses reasoning_format instead of reasoning_effort alone
-				requestOptions.reasoning_format = "parsed";
-				// Keep reasoning_effort for Groq
-			}
-			break;
-
-		case "anthropic":
-			// Anthropic's OpenAI compatibility has its own quirks
-			// But thinking content isn't available via OpenAI compat layer
-			break;
-
-		case "openrouter":
-			// OpenRouter uses a unified reasoning parameter format
-			if (api === "completions" && supportsReasoning && requestOptions.reasoning_effort) {
-				// Convert reasoning_effort to OpenRouter's reasoning format
-				requestOptions.reasoning = {
-					effort:
-						requestOptions.reasoning_effort === "low"
-							? "low"
-							: requestOptions.reasoning_effort === "minimal"
-								? "low"
-								: requestOptions.reasoning_effort === "medium"
-									? "medium"
-									: "high",
-				};
-				delete requestOptions.reasoning_effort;
-			}
-			break;
-
-		default:
-			// OpenAI and others use standard format
-			break;
-	}
-
-	return requestOptions;
-}
-
-async function checkReasoningSupport(
-	client: OpenAI,
-	model: string,
-	api: "completions" | "responses",
-	baseURL?: string,
-	signal?: AbortSignal,
-): Promise<boolean> {
-	// Check if already aborted
-	if (signal?.aborted) {
-		throw new Error("Interrupted");
-	}
-
-	// Check cache first
-	const cacheKey = model;
-	const cached = modelReasoningSupport.get(cacheKey);
-	if (cached && cached[api] !== undefined) {
-		return cached[api]!;
-	}
-
-	let supportsReasoning = false;
-	const provider = detectProvider(baseURL);
-
-	if (api === "responses") {
-		// Try a minimal request with reasoning parameter for Responses API
-		try {
-			const testRequest: any = {
-				model,
-				input: "test",
-				max_output_tokens: 1024,
-				reasoning: {
-					effort: "low", // Use low instead of minimal to ensure we get summaries
-				},
-			};
-			await client.responses.create(testRequest, { signal });
-			supportsReasoning = true;
-		} catch (error) {
-			supportsReasoning = false;
-		}
-	} else {
-		// For Chat Completions API, try with reasoning parameter
-		try {
-			const testRequest: any = {
-				model,
-				messages: [{ role: "user", content: "test" }],
-				max_completion_tokens: 1024,
-			};
-
-			// Add provider-specific reasoning parameters
-			if (provider === "gemini") {
-				// Gemini uses extra_body for thinking
-				testRequest.extra_body = {
-					google: {
-						thinking_config: {
-							thinking_budget: 100, // Minimum viable budget for test
-							include_thoughts: true,
-						},
-					},
-				};
-			} else if (provider === "groq") {
-				// Groq uses both reasoning_format and reasoning_effort
-				testRequest.reasoning_format = "parsed";
-				testRequest.reasoning_effort = "low";
-			} else {
-				// Others use reasoning_effort
-				testRequest.reasoning_effort = "minimal";
-			}
-
-			await client.chat.completions.create(testRequest, { signal });
-			supportsReasoning = true;
-		} catch (error) {
-			supportsReasoning = false;
+					// Add text blocks for documents with extracted text
+					else if (attachment.type === "document" && attachment.extractedText) {
+						content.push({
+							type: "text",
+							text: `\n\n[Document: ${attachment.fileName}]\n${attachment.extractedText}`,
+							isDocument: true,
+						} as TextContent);
 					}
 				}
 
-	// Update cache
-	const existing = modelReasoningSupport.get(cacheKey) || {};
-	existing[api] = supportsReasoning;
-	modelReasoningSupport.set(cacheKey, existing);
-
-	return supportsReasoning;
-}
-
-export async function callModelResponsesApi(
-	client: OpenAI,
-	model: string,
-	messages: any[],
-	signal?: AbortSignal,
-	eventReceiver?: AgentEventReceiver,
-	supportsReasoning?: boolean,
-	baseURL?: string,
-): Promise<void> {
-	let conversationDone = false;
-
-	while (!conversationDone) {
-		// Check if we've been interrupted
-		if (signal?.aborted) {
-			throw new Error("Interrupted");
+				return { ...rest, content } as Message;
 			}
-
-		// Build request options
-		let requestOptions: any = {
-			model,
-			input: messages,
-			tools: toolsForResponses as any,
-			tool_choice: "auto",
-			parallel_tool_calls: true,
-			max_output_tokens: 2000, // TODO make configurable
-			...(supportsReasoning && {
-				reasoning: {
-					effort: "minimal", // Use minimal effort for responses API
-					summary: "detailed", // Request detailed reasoning summaries
-				},
-			}),
-		};
-
-		// Apply provider-specific adjustments
-		requestOptions = adjustRequestForProvider(requestOptions, "responses", baseURL, supportsReasoning);
-
-		const response = await client.responses.create(requestOptions, { signal });
-
-		// Report token usage if available (responses API format)
-		if (response.usage) {
-			const usage = response.usage;
-			eventReceiver?.on({
-				type: "token_usage",
-				inputTokens: usage.input_tokens || 0,
-				outputTokens: usage.output_tokens || 0,
-				totalTokens: usage.total_tokens || 0,
-				cacheReadTokens: usage.input_tokens_details?.cached_tokens || 0,
-				cacheWriteTokens: 0, // Not available in API
-				reasoningTokens: usage.output_tokens_details?.reasoning_tokens || 0,
+			return m as Message;
 		});
-		}
-
-		const output = response.output;
-		if (!output) break;
-
-		for (const item of output) {
-			// gpt-oss vLLM quirk: need to remove type from "message" events
-			if (item.id === "message") {
-				const { type, ...message } = item;
-				messages.push(item);
-			} else {
-				messages.push(item);
-			}
-
-			switch (item.type) {
-				case "reasoning": {
-					// Handle both content (o1/o3) and summary (gpt-5) formats
-					const reasoningItems = item.content || item.summary || [];
-					for (const content of reasoningItems) {
-						if (content.type === "reasoning_text" || content.type === "summary_text") {
-							await eventReceiver?.on({ type: "reasoning", text: content.text });
-						}
-					}
-					break;
-				}
-
-				case "message": {
-					for (const content of item.content || []) {
-						if (content.type === "output_text") {
-							await eventReceiver?.on({ type: "assistant_message", text: content.text });
-						} else if (content.type === "refusal") {
-							await eventReceiver?.on({ type: "error", message: `Refusal: ${content.refusal}` });
-						}
-						conversationDone = true;
-					}
-					break;
-				}
-
-				case "function_call": {
-					if (signal?.aborted) {
-						throw new Error("Interrupted");
-					}
-
-					try {
-						await eventReceiver?.on({
-							type: "tool_call",
-							toolCallId: item.call_id || "",
-							name: item.name,
-							args: item.arguments,
-						});
-						const result = await executeTool(item.name, item.arguments, signal);
-						await eventReceiver?.on({
-							type: "tool_result",
-							toolCallId: item.call_id || "",
-							result,
-							isError: false,
-						});
-
-						// Add tool result to messages
-						const toolResultMsg = {
-							type: "function_call_output",
-							call_id: item.call_id,
-							output: result,
-						} as ResponseFunctionToolCallOutputItem;
-						messages.push(toolResultMsg);
-					} catch (e: any) {
-						await eventReceiver?.on({
-							type: "tool_result",
-							toolCallId: item.call_id || "",
-							result: e.message,
-							isError: true,
-						});
-						const errorMsg = {
-							type: "function_call_output",
-							call_id: item.id,
-							output: e.message,
-							isError: true,
-						};
-						messages.push(errorMsg);
-					}
-					break;
-				}
-
-				default: {
-					eventReceiver?.on({ type: "error", message: `Unknown output type in LLM response: ${item.type}` });
-					break;
-				}
-			}
-		}
-	}
 }
 
-export async function callModelChatCompletionsApi(
-	client: OpenAI,
-	model: string,
-	messages: any[],
-	signal?: AbortSignal,
-	eventReceiver?: AgentEventReceiver,
-	supportsReasoning?: boolean,
-	baseURL?: string,
-): Promise<void> {
-	let assistantResponded = false;
-
-	while (!assistantResponded) {
-		if (signal?.aborted) {
-			throw new Error("Interrupted");
-		}
-
-		// Build request options
-		let requestOptions: any = {
-			model,
-			messages,
-			tools: toolsForChat,
-			tool_choice: "auto",
-			max_completion_tokens: 2000, // TODO make configurable
-			...(supportsReasoning && {
-				reasoning_effort: "low", // Use low effort for completions API
-			}),
-		};
-
-		// Apply provider-specific adjustments
-		requestOptions = adjustRequestForProvider(requestOptions, "completions", baseURL, supportsReasoning);
-
-		const response = await client.chat.completions.create(requestOptions, { signal });
-
-		const message = response.choices[0].message;
-
-		// Report token usage if available
-		if (response.usage) {
-			const usage = response.usage;
-			await eventReceiver?.on({
-				type: "token_usage",
-				inputTokens: usage.prompt_tokens || 0,
-				outputTokens: usage.completion_tokens || 0,
-				totalTokens: usage.total_tokens || 0,
-				cacheReadTokens: usage.prompt_tokens_details?.cached_tokens || 0,
-				cacheWriteTokens: 0, // Not available in API
-				reasoningTokens: usage.completion_tokens_details?.reasoning_tokens || 0,
-			});
-		}
-
-		if (message.tool_calls && message.tool_calls.length > 0) {
-			// Add assistant message with tool calls to history
-			const assistantMsg: any = {
-				role: "assistant",
-				content: message.content || null,
-				tool_calls: message.tool_calls,
-			};
-			messages.push(assistantMsg);
-
-			// Display and execute each tool call
-			for (const toolCall of message.tool_calls) {
-				// Check if interrupted before executing tool
-				if (signal?.aborted) {
-					throw new Error("Interrupted");
-				}
-
-				try {
-					const funcName = toolCall.type === "function" ? toolCall.function.name : toolCall.custom.name;
-					const funcArgs = toolCall.type === "function" ? toolCall.function.arguments : toolCall.custom.input;
-
-					await eventReceiver?.on({ type: "tool_call", toolCallId: toolCall.id, name: funcName, args: funcArgs });
-					const result = await executeTool(funcName, funcArgs, signal);
-					await eventReceiver?.on({ type: "tool_result", toolCallId: toolCall.id, result, isError: false });
-
-					// Add tool result to messages
-					const toolMsg = {
-						role: "tool",
-						tool_call_id: toolCall.id,
-						content: result,
-					};
-					messages.push(toolMsg);
-				} catch (e: any) {
-					eventReceiver?.on({ type: "tool_result", toolCallId: toolCall.id, result: e.message, isError: true });
-					const errorMsg = {
-						role: "tool",
-						tool_call_id: toolCall.id,
-						content: e.message,
-					};
-					messages.push(errorMsg);
-				}
-			}
-		} else if (message.content) {
-			// Parse provider-specific reasoning from message
-			const { cleanContent, reasoningTexts } = parseReasoningFromMessage(message, baseURL);
-
-			// Emit reasoning events if any
-			for (const reasoning of reasoningTexts) {
-				await eventReceiver?.on({ type: "reasoning", text: reasoning });
-			}
-
-			// Emit the cleaned assistant message
-			await eventReceiver?.on({ type: "assistant_message", text: cleanContent });
-			const finalMsg = { role: "assistant", content: cleanContent };
-			messages.push(finalMsg);
-			assistantResponded = true;
-		}
-	}
+export interface AgentOptions {
+	initialState?: Partial<AgentState>;
+	transport: AgentTransport;
+	// Transform app messages to LLM-compatible messages before sending to transport
+	messageTransformer?: (messages: AppMessage[]) => Message[] | Promise<Message[]>;
 }
 
 export class Agent {
-	private client: OpenAI;
-	public readonly config: AgentConfig;
-	private messages: any[] = [];
-	private renderer?: AgentEventReceiver;
-	private sessionManager?: SessionManager;
-	private comboReceiver: AgentEventReceiver;
-	private abortController: AbortController | null = null;
-	private supportsReasoning: boolean | null = null;
-
-	constructor(config: AgentConfig, renderer?: AgentEventReceiver, sessionManager?: SessionManager) {
-		this.config = config;
-		this.client = new OpenAI({
-			apiKey: config.apiKey,
-			baseURL: config.baseURL,
-		});
-
-		// Use provided renderer or default to console
-		this.renderer = renderer;
-		this.sessionManager = sessionManager;
-
-		this.comboReceiver = {
-			on: async (event: AgentEvent): Promise<void> => {
-				await this.renderer?.on(event);
-				await this.sessionManager?.on(event);
-			},
+	private _state: AgentState = {
+		systemPrompt: "",
+		model: getModel("google", "gemini-2.5-flash-lite-preview-06-17"),
+		thinkingLevel: "off",
+		tools: [],
+		messages: [],
+		isStreaming: false,
+		streamMessage: null,
+		pendingToolCalls: new Set<string>(),
+		error: undefined,
 	};
+	private listeners = new Set<(e: AgentEvent) => void>();
+	private abortController?: AbortController;
+	private transport: AgentTransport;
+	private messageTransformer: (messages: AppMessage[]) => Message[] | Promise<Message[]>;
+	private messageQueue: Array<QueuedMessage<AppMessage>> = [];
 
-		// Initialize with system prompt if provided
-		if (config.systemPrompt) {
-			this.messages.push({
-				role: "developer",
-				content: config.systemPrompt,
+	constructor(opts: AgentOptions) {
+		this._state = { ...this._state, ...opts.initialState };
+		this.transport = opts.transport;
+		this.messageTransformer = opts.messageTransformer || defaultMessageTransformer;
+	}
+
+	get state(): AgentState {
+		return this._state;
+	}
+
+	subscribe(fn: (e: AgentEvent) => void): () => void {
+		this.listeners.add(fn);
+		fn({ type: "state-update", state: this._state });
+		return () => this.listeners.delete(fn);
+	}
+
+	// State mutators
+	setSystemPrompt(v: string) {
+		this.patch({ systemPrompt: v });
+	}
+
+	setModel(m: typeof this._state.model) {
+		this.patch({ model: m });
+	}
+
+	setThinkingLevel(l: ThinkingLevel) {
+		this.patch({ thinkingLevel: l });
+	}
+
+	setTools(t: typeof this._state.tools) {
+		this.patch({ tools: t });
+	}
+
+	replaceMessages(ms: AppMessage[]) {
+		this.patch({ messages: ms.slice() });
+	}
+
+	appendMessage(m: AppMessage) {
+		this.patch({ messages: [...this._state.messages, m] });
+	}
+
+	async queueMessage(m: AppMessage) {
+		// Transform message and queue it for injection at next turn
+		const transformed = await this.messageTransformer([m]);
+		this.messageQueue.push({
+			original: m,
+			llm: transformed[0], // undefined if filtered out
 		});
 	}
 
-		// Start session logging if we have a session manager
-		if (sessionManager) {
-			sessionManager.startSession(this.config);
-
-			// Emit session_start event
-			this.comboReceiver.on({
-				type: "session_start",
-				sessionId: sessionManager.getSessionId(),
-				model: config.model,
-				api: config.api,
-				baseURL: config.baseURL,
-				systemPrompt: config.systemPrompt,
-			});
-		}
+	clearMessages() {
+		this.patch({ messages: [] });
 	}
 
-	async ask(userMessage: string): Promise<void> {
-		// Render user message through the event system
-		this.comboReceiver.on({ type: "user_message", text: userMessage });
-
-		// Add user message
-		const userMsg = { role: "user", content: userMessage };
-		this.messages.push(userMsg);
-
-		// Create a new AbortController for this chat session
-		this.abortController = new AbortController();
-
-		try {
-			await this.comboReceiver.on({ type: "assistant_start" });
-
-			// Check reasoning support only once per agent instance
-			if (this.supportsReasoning === null) {
-				this.supportsReasoning = await checkReasoningSupport(
-					this.client,
-					this.config.model,
-					this.config.api,
-					this.config.baseURL,
-					this.abortController.signal,
-				);
-			}
-
-			if (this.config.api === "responses") {
-				await callModelResponsesApi(
-					this.client,
-					this.config.model,
-					this.messages,
-					this.abortController.signal,
-					this.comboReceiver,
-					this.supportsReasoning,
-					this.config.baseURL,
-				);
-			} else {
-				await callModelChatCompletionsApi(
-					this.client,
-					this.config.model,
-					this.messages,
-					this.abortController.signal,
-					this.comboReceiver,
-					this.supportsReasoning,
-					this.config.baseURL,
-				);
-			}
-		} catch (e) {
-			// Check if this was an interruption by checking the abort signal
-			if (this.abortController.signal.aborted) {
-				// Emit interrupted event so UI can clean up properly
-				await this.comboReceiver?.on({ type: "interrupted" });
-				return;
-			}
-			throw e;
-		} finally {
-			this.abortController = null;
-		}
-	}
-
-	interrupt(): void {
+	abort() {
 		this.abortController?.abort();
 	}
 
-	setEvents(events: AgentEvent[]): void {
-		// Reconstruct messages from events based on API type
-		this.messages = [];
-
-		if (this.config.api === "responses") {
-			// Responses API format
-			if (this.config.systemPrompt) {
-				this.messages.push({
-					role: "developer",
-					content: this.config.systemPrompt,
-				});
+	async prompt(input: string, attachments?: Attachment[]) {
+		const model = this._state.model;
+		if (!model) {
+			throw new Error("No model configured");
 		}
 
-			for (const event of events) {
-				switch (event.type) {
-					case "user_message":
-						this.messages.push({
+		// Build user message with attachments
+		const content: Array<TextContent | ImageContent> = [{ type: "text", text: input }];
+		if (attachments?.length) {
+			for (const a of attachments) {
+				if (a.type === "image") {
+					content.push({ type: "image", data: a.content, mimeType: a.mimeType });
+				} else if (a.type === "document" && a.extractedText) {
+					content.push({
+						type: "text",
+						text: `\n\n[Document: ${a.fileName}]\n${a.extractedText}`,
+						isDocument: true,
+					} as TextContent);
+				}
+			}
+		}
+
+		const userMessage: AppMessage = {
 			role: "user",
-							content: [{ type: "input_text", text: event.text }],
-						});
-						break;
+			content,
+			attachments: attachments?.length ? attachments : undefined,
+		};
 
-					case "reasoning":
-						// Add reasoning message
-						this.messages.push({
-							type: "reasoning",
-							content: [{ type: "reasoning_text", text: event.text }],
-						});
-						break;
+		this.abortController = new AbortController();
+		this.patch({ isStreaming: true, streamMessage: null, error: undefined });
+		this.emit({ type: "started" });
 
-					case "tool_call":
-						// Add function call
-						this.messages.push({
-							type: "function_call",
-							id: event.toolCallId,
-							name: event.name,
-							arguments: event.args,
-						});
-						break;
+		const reasoning =
+			this._state.thinkingLevel === "off"
+				? undefined
+				: this._state.thinkingLevel === "minimal"
+					? "low"
+					: this._state.thinkingLevel;
 
-					case "tool_result":
-						// Add function result
-						this.messages.push({
-							type: "function_call_output",
-							call_id: event.toolCallId,
-							output: event.result,
-						});
-						break;
-
-					case "assistant_message":
-						// Add final message
-						this.messages.push({
-							type: "message",
-							content: [{ type: "output_text", text: event.text }],
-						});
-						break;
-				}
-			}
-		} else {
-			// Chat Completions API format
-			if (this.config.systemPrompt) {
-				this.messages.push({ role: "system", content: this.config.systemPrompt });
-			}
-
-			// Track tool calls in progress
-			let pendingToolCalls: any[] = [];
-
-			for (const event of events) {
-				switch (event.type) {
-					case "user_message":
-						this.messages.push({ role: "user", content: event.text });
-						break;
-
-					case "assistant_start":
-						// Reset pending tool calls for new assistant response
-						pendingToolCalls = [];
-						break;
-
-					case "tool_call":
-						// Accumulate tool calls
-						pendingToolCalls.push({
-							id: event.toolCallId,
-							type: "function",
-							function: {
-								name: event.name,
-								arguments: event.args,
+		const cfg = {
+			systemPrompt: this._state.systemPrompt,
+			tools: this._state.tools,
+			model,
+			reasoning,
+			getQueuedMessages: async <T>() => {
+				// Return queued messages (they'll be added to state via message_end event)
+				const queued = this.messageQueue.slice();
+				this.messageQueue = [];
+				return queued as QueuedMessage<T>[];
 			},
-						});
-						break;
+		};
 
-					case "tool_result":
-						// When we see the first tool result, add the assistant message with all tool calls
-						if (pendingToolCalls.length > 0) {
-							this.messages.push({
+		try {
+			let partial: Message | null = null;
+
+			// Transform app messages to LLM-compatible messages (initial set)
+			const llmMessages = await this.messageTransformer(this._state.messages);
+
+			for await (const ev of this.transport.run(
+				llmMessages,
+				userMessage as Message,
+				cfg,
+				this.abortController.signal,
+			)) {
+				switch (ev.type) {
+					case "message_start":
+					case "message_update": {
+						partial = ev.message;
+						this.patch({ streamMessage: ev.message });
+						break;
+					}
+					case "message_end": {
+						partial = null;
+						this.appendMessage(ev.message as AppMessage);
+						this.patch({ streamMessage: null });
+						break;
+					}
+					case "tool_execution_start": {
+						const s = new Set(this._state.pendingToolCalls);
+						s.add(ev.toolCallId);
+						this.patch({ pendingToolCalls: s });
+						break;
+					}
+					case "tool_execution_end": {
+						const s = new Set(this._state.pendingToolCalls);
+						s.delete(ev.toolCallId);
+						this.patch({ pendingToolCalls: s });
+						break;
+					}
+					case "agent_end": {
+						this.patch({ streamMessage: null });
+						break;
+					}
+				}
+			}
+
+			if (partial && partial.role === "assistant" && partial.content.length > 0) {
+				const onlyEmpty = !partial.content.some(
+					(c) =>
+						(c.type === "thinking" && c.thinking.trim().length > 0) ||
+						(c.type === "text" && c.text.trim().length > 0) ||
+						(c.type === "toolCall" && c.name.trim().length > 0),
+				);
+				if (!onlyEmpty) {
+					this.appendMessage(partial as AppMessage);
+				} else {
+					if (this.abortController?.signal.aborted) {
+						throw new Error("Request was aborted");
+					}
+				}
+			}
+		} catch (err: any) {
+			const msg: Message = {
 				role: "assistant",
-								content: null,
-								tool_calls: pendingToolCalls,
-							});
-							pendingToolCalls = [];
+				content: [{ type: "text", text: "" }],
+				api: model.api,
+				provider: model.provider,
+				model: model.id,
+				usage: {
+					input: 0,
+					output: 0,
+					cacheRead: 0,
+					cacheWrite: 0,
+					cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+				},
+				stopReason: this.abortController?.signal.aborted ? "aborted" : "error",
+				errorMessage: err?.message || String(err),
+			};
+			this.appendMessage(msg as AppMessage);
+			this.patch({ error: err?.message || String(err) });
+		} finally {
+			this.patch({ isStreaming: false, streamMessage: null, pendingToolCalls: new Set<string>() });
+			this.abortController = undefined;
+			this.emit({ type: "completed" });
+		}
 	}
-						// Add the tool result
-						this.messages.push({
-							role: "tool",
-							tool_call_id: event.toolCallId,
-							content: event.result,
-						});
-						break;
 
-					case "assistant_message":
-						// Final assistant response (no tool calls)
-						this.messages.push({ role: "assistant", content: event.text });
-						break;
+	private patch(p: Partial<AgentState>): void {
+		this._state = { ...this._state, ...p };
+		this.emit({ type: "state-update", state: this._state });
+	}
 
-					// Skip other event types (thinking, error, interrupted, token_usage)
-				}
-			}
+	private emit(e: AgentEvent) {
+		for (const listener of this.listeners) {
+			listener(e);
 		}
 	}
 }

packages/agent/src/index.ts

@@ -1,15 +1,22 @@
-// Main exports for pi-agent package
-
-export type { AgentConfig, AgentEvent, AgentEventReceiver } from "./agent.js";
-export { Agent } from "./agent.js";
-export type { ArgDef, ArgDefs, ParsedArgs } from "./args.js";
-// CLI utilities
-export { parseArgs, printHelp } from "./args.js";
-// CLI main function
-export { main } from "./main.js";
-// Renderers
-export { ConsoleRenderer } from "./renderers/console-renderer.js";
-export { JsonRenderer } from "./renderers/json-renderer.js";
-export { TuiRenderer } from "./renderers/tui-renderer.js";
-export type { SessionData, SessionEvent, SessionHeader } from "./session-manager.js";
-export { SessionManager } from "./session-manager.js";
+// Core Agent
+export { Agent, type AgentOptions } from "./agent.js";
+// Transports
+export {
+	type AgentRunConfig,
+	type AgentTransport,
+	AppTransport,
+	type AppTransportOptions,
+	ProviderTransport,
+	type ProviderTransportOptions,
+	type ProxyAssistantMessageEvent,
+} from "./transports/index.js";
+// Types
+export type {
+	AgentEvent,
+	AgentState,
+	AppMessage,
+	Attachment,
+	CustomMessages,
+	ThinkingLevel,
+	UserMessageWithAttachments,
+} from "./types.js";

packages/agent/src/transports/AppTransport.ts

@@ -0,0 +1,374 @@
+import type {
+	AgentContext,
+	AgentLoopConfig,
+	Api,
+	AssistantMessage,
+	AssistantMessageEvent,
+	Context,
+	Message,
+	Model,
+	SimpleStreamOptions,
+	ToolCall,
+	UserMessage,
+} from "@mariozechner/pi-ai";
+import { agentLoop } from "@mariozechner/pi-ai";
+import { AssistantMessageEventStream } from "@mariozechner/pi-ai/dist/utils/event-stream.js";
+import { parseStreamingJson } from "@mariozechner/pi-ai/dist/utils/json-parse.js";
+import type { ProxyAssistantMessageEvent } from "./proxy-types.js";
+import type { AgentRunConfig, AgentTransport } from "./types.js";
+
+/**
+ * Stream function that proxies through a server instead of calling providers directly.
+ * The server strips the partial field from delta events to reduce bandwidth.
+ * We reconstruct the partial message client-side.
+ */
+function streamSimpleProxy(
+	model: Model<any>,
+	context: Context,
+	options: SimpleStreamOptions & { authToken: string },
+	proxyUrl: string,
+): AssistantMessageEventStream {
+	const stream = new AssistantMessageEventStream();
+
+	(async () => {
+		// Initialize the partial message that we'll build up from events
+		const partial: AssistantMessage = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [],
+			api: model.api,
+			provider: model.provider,
+			model: model.id,
+			usage: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+			},
+		};
+
+		let reader: ReadableStreamDefaultReader<Uint8Array> | undefined;
+
+		// Set up abort handler to cancel the reader
+		const abortHandler = () => {
+			if (reader) {
+				reader.cancel("Request aborted by user").catch(() => {});
+			}
+		};
+
+		if (options.signal) {
+			options.signal.addEventListener("abort", abortHandler);
+		}
+
+		try {
+			const response = await fetch(`${proxyUrl}/api/stream`, {
+				method: "POST",
+				headers: {
+					Authorization: `Bearer ${options.authToken}`,
+					"Content-Type": "application/json",
+				},
+				body: JSON.stringify({
+					model,
+					context,
+					options: {
+						temperature: options.temperature,
+						maxTokens: options.maxTokens,
+						reasoning: options.reasoning,
+						// Don't send apiKey or signal - those are added server-side
+					},
+				}),
+				signal: options.signal,
+			});
+
+			if (!response.ok) {
+				let errorMessage = `Proxy error: ${response.status} ${response.statusText}`;
+				try {
+					const errorData = (await response.json()) as { error?: string };
+					if (errorData.error) {
+						errorMessage = `Proxy error: ${errorData.error}`;
+					}
+				} catch {
+					// Couldn't parse error response, use default message
+				}
+				throw new Error(errorMessage);
+			}
+
+			// Parse SSE stream
+			reader = response.body!.getReader();
+			const decoder = new TextDecoder();
+			let buffer = "";
+
+			while (true) {
+				const { done, value } = await reader.read();
+				if (done) break;
+
+				// Check if aborted after reading
+				if (options.signal?.aborted) {
+					throw new Error("Request aborted by user");
+				}
+
+				buffer += decoder.decode(value, { stream: true });
+				const lines = buffer.split("\n");
+				buffer = lines.pop() || "";
+
+				for (const line of lines) {
+					if (line.startsWith("data: ")) {
+						const data = line.slice(6).trim();
+						if (data) {
+							const proxyEvent = JSON.parse(data) as ProxyAssistantMessageEvent;
+							let event: AssistantMessageEvent | undefined;
+
+							// Handle different event types
+							// Server sends events with partial for non-delta events,
+							// and without partial for delta events
+							switch (proxyEvent.type) {
+								case "start":
+									event = { type: "start", partial };
+									break;
+
+								case "text_start":
+									partial.content[proxyEvent.contentIndex] = {
+										type: "text",
+										text: "",
+									};
+									event = { type: "text_start", contentIndex: proxyEvent.contentIndex, partial };
+									break;
+
+								case "text_delta": {
+									const content = partial.content[proxyEvent.contentIndex];
+									if (content?.type === "text") {
+										content.text += proxyEvent.delta;
+										event = {
+											type: "text_delta",
+											contentIndex: proxyEvent.contentIndex,
+											delta: proxyEvent.delta,
+											partial,
+										};
+									} else {
+										throw new Error("Received text_delta for non-text content");
+									}
+									break;
+								}
+								case "text_end": {
+									const content = partial.content[proxyEvent.contentIndex];
+									if (content?.type === "text") {
+										content.textSignature = proxyEvent.contentSignature;
+										event = {
+											type: "text_end",
+											contentIndex: proxyEvent.contentIndex,
+											content: content.text,
+											partial,
+										};
+									} else {
+										throw new Error("Received text_end for non-text content");
+									}
+									break;
+								}
+
+								case "thinking_start":
+									partial.content[proxyEvent.contentIndex] = {
+										type: "thinking",
+										thinking: "",
+									};
+									event = { type: "thinking_start", contentIndex: proxyEvent.contentIndex, partial };
+									break;
+
+								case "thinking_delta": {
+									const content = partial.content[proxyEvent.contentIndex];
+									if (content?.type === "thinking") {
+										content.thinking += proxyEvent.delta;
+										event = {
+											type: "thinking_delta",
+											contentIndex: proxyEvent.contentIndex,
+											delta: proxyEvent.delta,
+											partial,
+										};
+									} else {
+										throw new Error("Received thinking_delta for non-thinking content");
+									}
+									break;
+								}
+
+								case "thinking_end": {
+									const content = partial.content[proxyEvent.contentIndex];
+									if (content?.type === "thinking") {
+										content.thinkingSignature = proxyEvent.contentSignature;
+										event = {
+											type: "thinking_end",
+											contentIndex: proxyEvent.contentIndex,
+											content: content.thinking,
+											partial,
+										};
+									} else {
+										throw new Error("Received thinking_end for non-thinking content");
+									}
+									break;
+								}
+
+								case "toolcall_start":
+									partial.content[proxyEvent.contentIndex] = {
+										type: "toolCall",
+										id: proxyEvent.id,
+										name: proxyEvent.toolName,
+										arguments: {},
+										partialJson: "",
+									} satisfies ToolCall & { partialJson: string } as ToolCall;
+									event = { type: "toolcall_start", contentIndex: proxyEvent.contentIndex, partial };
+									break;
+
+								case "toolcall_delta": {
+									const content = partial.content[proxyEvent.contentIndex];
+									if (content?.type === "toolCall") {
+										(content as any).partialJson += proxyEvent.delta;
+										content.arguments = parseStreamingJson((content as any).partialJson) || {};
+										event = {
+											type: "toolcall_delta",
+											contentIndex: proxyEvent.contentIndex,
+											delta: proxyEvent.delta,
+											partial,
+										};
+										partial.content[proxyEvent.contentIndex] = { ...content }; // Trigger reactivity
+									} else {
+										throw new Error("Received toolcall_delta for non-toolCall content");
+									}
+									break;
+								}
+
+								case "toolcall_end": {
+									const content = partial.content[proxyEvent.contentIndex];
+									if (content?.type === "toolCall") {
+										delete (content as any).partialJson;
+										event = {
+											type: "toolcall_end",
+											contentIndex: proxyEvent.contentIndex,
+											toolCall: content,
+											partial,
+										};
+									}
+									break;
+								}
+
+								case "done":
+									partial.stopReason = proxyEvent.reason;
+									partial.usage = proxyEvent.usage;
+									event = { type: "done", reason: proxyEvent.reason, message: partial };
+									break;
+
+								case "error":
+									partial.stopReason = proxyEvent.reason;
+									partial.errorMessage = proxyEvent.errorMessage;
+									partial.usage = proxyEvent.usage;
+									event = { type: "error", reason: proxyEvent.reason, error: partial };
+									break;
+
+								default: {
+									// Exhaustive check
+									const _exhaustiveCheck: never = proxyEvent;
+									console.warn(`Unhandled event type: ${(proxyEvent as any).type}`);
+									break;
+								}
+							}
+
+							// Push the event to stream
+							if (event) {
+								stream.push(event);
+							} else {
+								throw new Error("Failed to create event from proxy event");
+							}
+						}
+					}
+				}
+			}
+
+			// Check if aborted after reading
+			if (options.signal?.aborted) {
+				throw new Error("Request aborted by user");
+			}
+
+			stream.end();
+		} catch (error) {
+			const errorMessage = error instanceof Error ? error.message : String(error);
+			partial.stopReason = options.signal?.aborted ? "aborted" : "error";
+			partial.errorMessage = errorMessage;
+			stream.push({
+				type: "error",
+				reason: partial.stopReason,
+				error: partial,
+			} satisfies AssistantMessageEvent);
+			stream.end();
+		} finally {
+			// Clean up abort handler
+			if (options.signal) {
+				options.signal.removeEventListener("abort", abortHandler);
+			}
+		}
+	})();
+
+	return stream;
+}
+
+export interface AppTransportOptions {
+	/**
+	 * Proxy server URL. The server manages user accounts and proxies requests to LLM providers.
+	 * Example: "https://genai.mariozechner.at"
+	 */
+	proxyUrl: string;
+
+	/**
+	 * Function to retrieve auth token for the proxy server.
+	 * The token is used for user authentication and authorization.
+	 */
+	getAuthToken: () => Promise<string> | string;
+}
+
+/**
+ * Transport that uses an app server with user authentication tokens.
+ * The server manages user accounts and proxies requests to LLM providers.
+ */
+export class AppTransport implements AgentTransport {
+	private options: AppTransportOptions;
+
+	constructor(options: AppTransportOptions) {
+		this.options = options;
+	}
+
+	async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
+		const authToken = await this.options.getAuthToken();
+		if (!authToken) {
+			throw new Error("Auth token is required for AppTransport");
+		}
+
+		// Use proxy - no local API key needed
+		const streamFn = <TApi extends Api>(model: Model<TApi>, context: Context, options?: SimpleStreamOptions) => {
+			return streamSimpleProxy(
+				model,
+				context,
+				{
+					...options,
+					authToken,
+				},
+				this.options.proxyUrl,
+			);
+		};
+
+		// Messages are already LLM-compatible (filtered by Agent)
+		const context: AgentContext = {
+			systemPrompt: cfg.systemPrompt,
+			messages,
+			tools: cfg.tools,
+		};
+
+		const pc: AgentLoopConfig = {
+			model: cfg.model,
+			reasoning: cfg.reasoning,
+			getQueuedMessages: cfg.getQueuedMessages,
+		};
+
+		// Yield events from the upstream agentLoop iterator
+		// Pass streamFn as the 5th parameter to use proxy
+		for await (const ev of agentLoop(userMessage as unknown as UserMessage, context, pc, signal, streamFn as any)) {
+			yield ev;
+		}
+	}
+}

packages/agent/src/transports/ProviderTransport.ts

@@ -0,0 +1,75 @@
+import {
+	type AgentContext,
+	type AgentLoopConfig,
+	agentLoop,
+	type Message,
+	type UserMessage,
+} from "@mariozechner/pi-ai";
+import type { AgentRunConfig, AgentTransport } from "./types.js";
+
+export interface ProviderTransportOptions {
+	/**
+	 * Function to retrieve API key for a given provider.
+	 * If not provided, transport will try to use environment variables.
+	 */
+	getApiKey?: (provider: string) => Promise<string | undefined> | string | undefined;
+
+	/**
+	 * Optional CORS proxy URL for browser environments.
+	 * If provided, all requests will be routed through this proxy.
+	 * Format: "https://proxy.example.com"
+	 */
+	corsProxyUrl?: string;
+}
+
+/**
+ * Transport that calls LLM providers directly.
+ * Optionally routes calls through a CORS proxy if configured.
+ */
+export class ProviderTransport implements AgentTransport {
+	private options: ProviderTransportOptions;
+
+	constructor(options: ProviderTransportOptions = {}) {
+		this.options = options;
+	}
+
+	async *run(messages: Message[], userMessage: Message, cfg: AgentRunConfig, signal?: AbortSignal) {
+		// Get API key
+		let apiKey: string | undefined;
+		if (this.options.getApiKey) {
+			apiKey = await this.options.getApiKey(cfg.model.provider);
+		}
+
+		if (!apiKey) {
+			throw new Error(`No API key found for provider: ${cfg.model.provider}`);
+		}
+
+		// Clone model and modify baseUrl if CORS proxy is enabled
+		let model = cfg.model;
+		if (this.options.corsProxyUrl && cfg.model.baseUrl) {
+			model = {
+				...cfg.model,
+				baseUrl: `${this.options.corsProxyUrl}/?url=${encodeURIComponent(cfg.model.baseUrl)}`,
+			};
+		}
+
+		// Messages are already LLM-compatible (filtered by Agent)
+		const context: AgentContext = {
+			systemPrompt: cfg.systemPrompt,
+			messages,
+			tools: cfg.tools,
+		};
+
+		const pc: AgentLoopConfig = {
+			model,
+			reasoning: cfg.reasoning,
+			apiKey,
+			getQueuedMessages: cfg.getQueuedMessages,
+		};
+
+		// Yield events from agentLoop
+		for await (const ev of agentLoop(userMessage as unknown as UserMessage, context, pc, signal)) {
+			yield ev;
+		}
+	}
+}

packages/agent/src/transports/index.ts

@@ -0,0 +1,4 @@
+export { AppTransport, type AppTransportOptions } from "./AppTransport.js";
+export { ProviderTransport, type ProviderTransportOptions } from "./ProviderTransport.js";
+export type { ProxyAssistantMessageEvent } from "./proxy-types.js";
+export type { AgentRunConfig, AgentTransport } from "./types.js";

packages/agent/src/transports/proxy-types.ts

@@ -0,0 +1,20 @@
+import type { StopReason, Usage } from "@mariozechner/pi-ai";
+
+/**
+ * Event types emitted by the proxy server.
+ * The server strips the `partial` field from delta events to reduce bandwidth.
+ * Clients reconstruct the partial message from these events.
+ */
+export type ProxyAssistantMessageEvent =
+	| { type: "start" }
+	| { type: "text_start"; contentIndex: number }
+	| { type: "text_delta"; contentIndex: number; delta: string }
+	| { type: "text_end"; contentIndex: number; contentSignature?: string }
+	| { type: "thinking_start"; contentIndex: number }
+	| { type: "thinking_delta"; contentIndex: number; delta: string }
+	| { type: "thinking_end"; contentIndex: number; contentSignature?: string }
+	| { type: "toolcall_start"; contentIndex: number; id: string; toolName: string }
+	| { type: "toolcall_delta"; contentIndex: number; delta: string }
+	| { type: "toolcall_end"; contentIndex: number }
+	| { type: "done"; reason: Extract<StopReason, "stop" | "length" | "toolUse">; usage: Usage }
+	| { type: "error"; reason: Extract<StopReason, "aborted" | "error">; errorMessage: string; usage: Usage };

packages/agent/src/transports/types.ts

@@ -0,0 +1,28 @@
+import type { AgentEvent, AgentTool, Message, Model, QueuedMessage } from "@mariozechner/pi-ai";
+
+/**
+ * The minimal configuration needed to run an agent turn.
+ */
+export interface AgentRunConfig {
+	systemPrompt: string;
+	tools: AgentTool<any>[];
+	model: Model<any>;
+	reasoning?: "low" | "medium" | "high";
+	getQueuedMessages?: <T>() => Promise<QueuedMessage<T>[]>;
+}
+
+/**
+ * Transport interface for executing agent turns.
+ * Transports handle the communication with LLM providers,
+ * abstracting away the details of API calls, proxies, etc.
+ *
+ * Events yielded must match the @mariozechner/pi-ai AgentEvent types.
+ */
+export interface AgentTransport {
+	run(
+		messages: Message[],
+		userMessage: Message,
+		config: AgentRunConfig,
+		signal?: AbortSignal,
+	): AsyncIterable<AgentEvent>;
+}

packages/agent/src/types.ts

@@ -0,0 +1,75 @@
+import type { AgentTool, AssistantMessage, Message, Model, UserMessage } from "@mariozechner/pi-ai";
+
+/**
+ * Attachment type definition.
+ * Processing is done by consumers (e.g., document extraction in web-ui).
+ */
+export interface Attachment {
+	id: string;
+	type: "image" | "document";
+	fileName: string;
+	mimeType: string;
+	size: number;
+	content: string; // base64 encoded (without data URL prefix)
+	extractedText?: string; // For documents
+	preview?: string; // base64 image preview
+}
+
+/**
+ * Thinking/reasoning level for models that support it.
+ */
+export type ThinkingLevel = "off" | "minimal" | "low" | "medium" | "high";
+
+/**
+ * User message with optional attachments.
+ */
+export type UserMessageWithAttachments = UserMessage & { attachments?: Attachment[] };
+
+/**
+ * Extensible interface for custom app messages.
+ * Apps can extend via declaration merging:
+ *
+ * @example
+ * ```typescript
+ * declare module "@mariozechner/agent" {
+ *   interface CustomMessages {
+ *     artifact: ArtifactMessage;
+ *     notification: NotificationMessage;
+ *   }
+ * }
+ * ```
+ */
+export interface CustomMessages {
+	// Empty by default - apps extend via declaration merging
+}
+
+/**
+ * AppMessage: Union of LLM messages + attachments + custom messages.
+ * This abstraction allows apps to add custom message types while maintaining
+ * type safety and compatibility with the base LLM messages.
+ */
+export type AppMessage =
+	| AssistantMessage
+	| UserMessageWithAttachments
+	| Message // Includes ToolResultMessage
+	| CustomMessages[keyof CustomMessages];
+
+/**
+ * Agent state containing all configuration and conversation data.
+ */
+export interface AgentState {
+	systemPrompt: string;
+	model: Model<any>;
+	thinkingLevel: ThinkingLevel;
+	tools: AgentTool<any>[];
+	messages: AppMessage[]; // Can include attachments + custom message types
+	isStreaming: boolean;
+	streamMessage: Message | null;
+	pendingToolCalls: Set<string>;
+	error?: string;
+}
+
+/**
+ * Events emitted by the Agent for UI updates.
+ */
+export type AgentEvent = { type: "state-update"; state: AgentState } | { type: "started" } | { type: "completed" };

packages/agent/test/agent.test.ts

@@ -0,0 +1,140 @@
+import { getModel } from "@mariozechner/pi-ai";
+import { describe, expect, it } from "vitest";
+import { Agent, ProviderTransport } from "../src/index.js";
+
+describe("Agent", () => {
+	it("should create an agent instance with default state", () => {
+		const agent = new Agent({
+			transport: new ProviderTransport(),
+		});
+
+		expect(agent.state).toBeDefined();
+		expect(agent.state.systemPrompt).toBe("");
+		expect(agent.state.model).toBeDefined();
+		expect(agent.state.thinkingLevel).toBe("off");
+		expect(agent.state.tools).toEqual([]);
+		expect(agent.state.messages).toEqual([]);
+		expect(agent.state.isStreaming).toBe(false);
+		expect(agent.state.streamMessage).toBe(null);
+		expect(agent.state.pendingToolCalls).toEqual(new Set());
+		expect(agent.state.error).toBeUndefined();
+	});
+
+	it("should create an agent instance with custom initial state", () => {
+		const customModel = getModel("openai", "gpt-4o-mini");
+		const agent = new Agent({
+			transport: new ProviderTransport(),
+			initialState: {
+				systemPrompt: "You are a helpful assistant.",
+				model: customModel,
+				thinkingLevel: "low",
+			},
+		});
+
+		expect(agent.state.systemPrompt).toBe("You are a helpful assistant.");
+		expect(agent.state.model).toBe(customModel);
+		expect(agent.state.thinkingLevel).toBe("low");
+	});
+
+	it("should subscribe to state updates", () => {
+		const agent = new Agent({
+			transport: new ProviderTransport(),
+		});
+
+		let updateCount = 0;
+		const unsubscribe = agent.subscribe((event) => {
+			if (event.type === "state-update") {
+				updateCount++;
+			}
+		});
+
+		// Initial state update on subscribe
+		expect(updateCount).toBe(1);
+
+		// Update state
+		agent.setSystemPrompt("Test prompt");
+		expect(updateCount).toBe(2);
+		expect(agent.state.systemPrompt).toBe("Test prompt");
+
+		// Unsubscribe should work
+		unsubscribe();
+		agent.setSystemPrompt("Another prompt");
+		expect(updateCount).toBe(2); // Should not increase
+	});
+
+	it("should update state with mutators", () => {
+		const agent = new Agent({
+			transport: new ProviderTransport(),
+		});
+
+		// Test setSystemPrompt
+		agent.setSystemPrompt("Custom prompt");
+		expect(agent.state.systemPrompt).toBe("Custom prompt");
+
+		// Test setModel
+		const newModel = getModel("google", "gemini-2.5-flash");
+		agent.setModel(newModel);
+		expect(agent.state.model).toBe(newModel);
+
+		// Test setThinkingLevel
+		agent.setThinkingLevel("high");
+		expect(agent.state.thinkingLevel).toBe("high");
+
+		// Test setTools
+		const tools = [{ name: "test", description: "test tool" } as any];
+		agent.setTools(tools);
+		expect(agent.state.tools).toBe(tools);
+
+		// Test replaceMessages
+		const messages = [{ role: "user" as const, content: "Hello" }];
+		agent.replaceMessages(messages);
+		expect(agent.state.messages).toEqual(messages);
+		expect(agent.state.messages).not.toBe(messages); // Should be a copy
+
+		// Test appendMessage
+		const newMessage = { role: "assistant" as const, content: [{ type: "text" as const, text: "Hi" }] };
+		agent.appendMessage(newMessage as any);
+		expect(agent.state.messages).toHaveLength(2);
+		expect(agent.state.messages[1]).toBe(newMessage);
+
+		// Test clearMessages
+		agent.clearMessages();
+		expect(agent.state.messages).toEqual([]);
+	});
+
+	it("should support message queueing", async () => {
+		const agent = new Agent({
+			transport: new ProviderTransport(),
+		});
+
+		const message = { role: "user" as const, content: "Queued message" };
+		await agent.queueMessage(message);
+
+		// The message is queued but not yet in state.messages
+		expect(agent.state.messages).not.toContainEqual(message);
+	});
+
+	it("should handle abort controller", () => {
+		const agent = new Agent({
+			transport: new ProviderTransport(),
+		});
+
+		// Should not throw even if nothing is running
+		expect(() => agent.abort()).not.toThrow();
+	});
+});
+
+describe("ProviderTransport", () => {
+	it("should create a provider transport instance", () => {
+		const transport = new ProviderTransport();
+		expect(transport).toBeDefined();
+	});
+
+	it("should create a provider transport with options", () => {
+		const transport = new ProviderTransport({
+			getApiKey: async (provider) => `test-key-${provider}`,
+			corsProxyUrl: "https://proxy.example.com",
+		});
+		expect(transport).toBeDefined();
+	});
+});

packages/agent/test/e2e.test.ts

@@ -0,0 +1,402 @@
+import type { Model } from "@mariozechner/pi-ai";
+import { calculateTool, getModel } from "@mariozechner/pi-ai";
+import { describe, expect, it } from "vitest";
+import { Agent, ProviderTransport } from "../src/index.js";
+
+async function basicPrompt(model: Model<any>) {
+	const agent = new Agent({
+		initialState: {
+			systemPrompt: "You are a helpful assistant. Keep your responses concise.",
+			model,
+			thinkingLevel: "off",
+			tools: [],
+		},
+		transport: new ProviderTransport({
+			getApiKey: async (provider) => {
+				// Map provider names to env var names
+				const envVarMap: Record<string, string> = {
+					google: "GEMINI_API_KEY",
+					openai: "OPENAI_API_KEY",
+					anthropic: "ANTHROPIC_API_KEY",
+					xai: "XAI_API_KEY",
+					groq: "GROQ_API_KEY",
+					cerebras: "CEREBRAS_API_KEY",
+					zai: "ZAI_API_KEY",
+				};
+				const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
+				return process.env[envVar];
+			},
+		}),
+	});
+
+	await agent.prompt("What is 2+2? Answer with just the number.");
+
+	expect(agent.state.isStreaming).toBe(false);
+	expect(agent.state.messages.length).toBe(2);
+	expect(agent.state.messages[0].role).toBe("user");
+	expect(agent.state.messages[1].role).toBe("assistant");
+
+	const assistantMessage = agent.state.messages[1];
+	if (assistantMessage.role !== "assistant") throw new Error("Expected assistant message");
+	expect(assistantMessage.content.length).toBeGreaterThan(0);
+
+	const textContent = assistantMessage.content.find((c) => c.type === "text");
+	expect(textContent).toBeDefined();
+	if (textContent?.type !== "text") throw new Error("Expected text content");
+	expect(textContent.text).toContain("4");
+}
+
+async function toolExecution(model: Model<any>) {
+	const agent = new Agent({
+		initialState: {
+			systemPrompt: "You are a helpful assistant. Always use the calculator tool for math.",
+			model,
+			thinkingLevel: "off",
+			tools: [calculateTool],
+		},
+		transport: new ProviderTransport({
+			getApiKey: async (provider) => {
+				// Map provider names to env var names
+				const envVarMap: Record<string, string> = {
+					google: "GEMINI_API_KEY",
+					openai: "OPENAI_API_KEY",
+					anthropic: "ANTHROPIC_API_KEY",
+					xai: "XAI_API_KEY",
+					groq: "GROQ_API_KEY",
+					cerebras: "CEREBRAS_API_KEY",
+					zai: "ZAI_API_KEY",
+				};
+				const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
+				return process.env[envVar];
+			},
+		}),
+	});
+
+	await agent.prompt("Calculate 123 * 456 using the calculator tool.");
+
+	expect(agent.state.isStreaming).toBe(false);
+	expect(agent.state.messages.length).toBeGreaterThanOrEqual(3);
+
+	const toolResultMsg = agent.state.messages.find((m) => m.role === "toolResult");
+	expect(toolResultMsg).toBeDefined();
+	if (toolResultMsg?.role !== "toolResult") throw new Error("Expected tool result message");
+	expect(toolResultMsg.output).toBeDefined();
+
+	const expectedResult = 123 * 456;
+	expect(toolResultMsg.output).toContain(String(expectedResult));
+
+	const finalMessage = agent.state.messages[agent.state.messages.length - 1];
+	if (finalMessage.role !== "assistant") throw new Error("Expected final assistant message");
+	const finalText = finalMessage.content.find((c) => c.type === "text");
+	expect(finalText).toBeDefined();
+	if (finalText?.type !== "text") throw new Error("Expected text content");
+	// Check for number with or without comma formatting
+	const hasNumber =
+		finalText.text.includes(String(expectedResult)) ||
+		finalText.text.includes("56,088") ||
+		finalText.text.includes("56088");
+	expect(hasNumber).toBe(true);
+}
+
+async function abortExecution(model: Model<any>) {
+	const agent = new Agent({
+		initialState: {
+			systemPrompt: "You are a helpful assistant.",
+			model,
+			thinkingLevel: "off",
+			tools: [calculateTool],
+		},
+		transport: new ProviderTransport({
+			getApiKey: async (provider) => {
+				// Map provider names to env var names
+				const envVarMap: Record<string, string> = {
+					google: "GEMINI_API_KEY",
+					openai: "OPENAI_API_KEY",
+					anthropic: "ANTHROPIC_API_KEY",
+					xai: "XAI_API_KEY",
+					groq: "GROQ_API_KEY",
+					cerebras: "CEREBRAS_API_KEY",
+					zai: "ZAI_API_KEY",
+				};
+				const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
+				return process.env[envVar];
+			},
+		}),
+	});
+
+	const promptPromise = agent.prompt("Calculate 100 * 200, then 300 * 400, then sum the results.");
+
+	setTimeout(() => {
+		agent.abort();
+	}, 100);
+
+	await promptPromise;
+
+	expect(agent.state.isStreaming).toBe(false);
+	expect(agent.state.messages.length).toBeGreaterThanOrEqual(2);
+
+	const lastMessage = agent.state.messages[agent.state.messages.length - 1];
+	if (lastMessage.role !== "assistant") throw new Error("Expected assistant message");
+	expect(lastMessage.stopReason).toBe("aborted");
+	expect(lastMessage.errorMessage).toBeDefined();
+}
+
+async function stateUpdates(model: Model<any>) {
+	const agent = new Agent({
+		initialState: {
+			systemPrompt: "You are a helpful assistant.",
+			model,
+			thinkingLevel: "off",
+			tools: [],
+		},
+		transport: new ProviderTransport({
+			getApiKey: async (provider) => {
+				// Map provider names to env var names
+				const envVarMap: Record<string, string> = {
+					google: "GEMINI_API_KEY",
+					openai: "OPENAI_API_KEY",
+					anthropic: "ANTHROPIC_API_KEY",
+					xai: "XAI_API_KEY",
+					groq: "GROQ_API_KEY",
+					cerebras: "CEREBRAS_API_KEY",
+					zai: "ZAI_API_KEY",
+				};
+				const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
+				return process.env[envVar];
+			},
+		}),
+	});
+
+	const stateSnapshots: Array<{ isStreaming: boolean; messageCount: number; hasStreamMessage: boolean }> = [];
+
+	agent.subscribe((event) => {
+		if (event.type === "state-update") {
+			stateSnapshots.push({
+				isStreaming: event.state.isStreaming,
+				messageCount: event.state.messages.length,
+				hasStreamMessage: event.state.streamMessage !== null,
+			});
+		}
+	});
+
+	await agent.prompt("Count from 1 to 5.");
+
+	const streamingStates = stateSnapshots.filter((s) => s.isStreaming);
+	const nonStreamingStates = stateSnapshots.filter((s) => !s.isStreaming);
+
+	expect(streamingStates.length).toBeGreaterThan(0);
+	expect(nonStreamingStates.length).toBeGreaterThan(0);
+
+	const finalState = stateSnapshots[stateSnapshots.length - 1];
+	expect(finalState.isStreaming).toBe(false);
+	expect(finalState.messageCount).toBe(2);
+}
+
+async function multiTurnConversation(model: Model<any>) {
+	const agent = new Agent({
+		initialState: {
+			systemPrompt: "You are a helpful assistant.",
+			model,
+			thinkingLevel: "off",
+			tools: [],
+		},
+		transport: new ProviderTransport({
+			getApiKey: async (provider) => {
+				// Map provider names to env var names
+				const envVarMap: Record<string, string> = {
+					google: "GEMINI_API_KEY",
+					openai: "OPENAI_API_KEY",
+					anthropic: "ANTHROPIC_API_KEY",
+					xai: "XAI_API_KEY",
+					groq: "GROQ_API_KEY",
+					cerebras: "CEREBRAS_API_KEY",
+					zai: "ZAI_API_KEY",
+				};
+				const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
+				return process.env[envVar];
+			},
+		}),
+	});
+
+	await agent.prompt("My name is Alice.");
+	expect(agent.state.messages.length).toBe(2);
+
+	await agent.prompt("What is my name?");
+	expect(agent.state.messages.length).toBe(4);
+
+	const lastMessage = agent.state.messages[3];
+	if (lastMessage.role !== "assistant") throw new Error("Expected assistant message");
+	const lastText = lastMessage.content.find((c) => c.type === "text");
+	if (lastText?.type !== "text") throw new Error("Expected text content");
+	expect(lastText.text.toLowerCase()).toContain("alice");
+}
+
+describe("Agent E2E Tests", () => {
+	describe.skipIf(!process.env.GEMINI_API_KEY)("Google Provider (gemini-2.5-flash)", () => {
+		const model = getModel("google", "gemini-2.5-flash");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+
+	describe.skipIf(!process.env.OPENAI_API_KEY)("OpenAI Provider (gpt-4o-mini)", () => {
+		const model = getModel("openai", "gpt-4o-mini");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+
+	describe.skipIf(!process.env.ANTHROPIC_API_KEY)("Anthropic Provider (claude-3-5-haiku-20241022)", () => {
+		const model = getModel("anthropic", "claude-3-5-haiku-20241022");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+
+	describe.skipIf(!process.env.XAI_API_KEY)("xAI Provider (grok-3)", () => {
+		const model = getModel("xai", "grok-3");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+
+	describe.skipIf(!process.env.GROQ_API_KEY)("Groq Provider (openai/gpt-oss-20b)", () => {
+		const model = getModel("groq", "openai/gpt-oss-20b");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+
+	describe.skipIf(!process.env.CEREBRAS_API_KEY)("Cerebras Provider (gpt-oss-120b)", () => {
+		const model = getModel("cerebras", "gpt-oss-120b");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+
+	describe.skipIf(!process.env.ZAI_API_KEY)("zAI Provider (glm-4.5-air)", () => {
+		const model = getModel("zai", "glm-4.5-air");
+
+		it("should handle basic text prompt", async () => {
+			await basicPrompt(model);
+		});
+
+		it("should execute tools correctly", async () => {
+			await toolExecution(model);
+		});
+
+		it("should handle abort during execution", async () => {
+			await abortExecution(model);
+		});
+
+		it("should emit state updates during streaming", async () => {
+			await stateUpdates(model);
+		});
+
+		it("should maintain context across multiple turns", async () => {
+			await multiTurnConversation(model);
+		});
+	});
+});

packages/agent/tsconfig.build.json

@@ -4,6 +4,6 @@
 		"outDir": "./dist",
 		"rootDir": "./src"
 	},
-	"include": ["src/**/*"],
-	"exclude": ["node_modules", "dist"]
+	"include": ["src/**/*.ts"],
+	"exclude": ["node_modules", "dist", "**/*.d.ts", "src/**/*.d.ts"]
 }

packages/agent/vitest.config.ts

@@ -0,0 +1,9 @@
+import { defineConfig } from "vitest/config";
+
+export default defineConfig({
+	test: {
+		globals: true,
+		environment: "node",
+		testTimeout: 10000, // 10 seconds
+	},
+});

packages/coding-agent/package.json

@@ -0,0 +1,52 @@
+{
+	"name": "@mariozechner/coding-agent",
+	"version": "0.5.44",
+	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
+	"type": "module",
+	"bin": {
+		"coding-agent": "dist/cli.js"
+	},
+	"main": "./dist/index.js",
+	"types": "./dist/index.d.ts",
+	"files": [
+		"dist"
+	],
+	"scripts": {
+		"clean": "rm -rf dist",
+		"build": "tsc -p tsconfig.build.json && chmod +x dist/cli.js",
+		"dev": "tsc -p tsconfig.build.json --watch --preserveWatchOutput",
+		"check": "tsc --noEmit",
+		"test": "vitest --run",
+		"prepublishOnly": "npm run clean && npm run build"
+	},
+	"dependencies": {
+		"@mariozechner/pi-agent": "^0.5.44",
+		"@mariozechner/pi-ai": "^0.5.44",
+		"@mariozechner/pi-tui": "^0.5.44",
+		"chalk": "^5.5.0",
+		"glob": "^11.0.3"
+	},
+	"devDependencies": {
+		"@types/node": "^24.3.0",
+		"typescript": "^5.7.3",
+		"vitest": "^3.2.4"
+	},
+	"keywords": [
+		"coding-agent",
+		"ai",
+		"llm",
+		"cli",
+		"tui",
+		"agent"
+	],
+	"author": "Mario Zechner",
+	"license": "MIT",
+	"repository": {
+		"type": "git",
+		"url": "git+https://github.com/badlogic/pi-mono.git",
+		"directory": "packages/coding-agent"
+	},
+	"engines": {
+		"node": ">=20.0.0"
+	}
+}

packages/coding-agent/src/cli.ts

@@ -0,0 +1,8 @@
+#!/usr/bin/env node
+
+import { main } from "./main.js";
+
+main(process.argv.slice(2)).catch((err) => {
+	console.error(err);
+	process.exit(1);
+});

packages/coding-agent/src/index.ts

@@ -0,0 +1,3 @@
+export { main } from "./main.js";
+export { SessionManager } from "./session-manager.js";
+export { bashTool, codingTools, editTool, readTool, writeTool } from "./tools/index.js";

packages/coding-agent/src/main.ts

@@ -0,0 +1,209 @@
+import { Agent, ProviderTransport } from "@mariozechner/pi-agent";
+import { getModel } from "@mariozechner/pi-ai";
+import chalk from "chalk";
+import { SessionManager } from "./session-manager.js";
+import { codingTools } from "./tools/index.js";
+
+interface Args {
+	provider?: string;
+	model?: string;
+	apiKey?: string;
+	systemPrompt?: string;
+	continue?: boolean;
+	help?: boolean;
+	messages: string[];
+}
+
+function parseArgs(args: string[]): Args {
+	const result: Args = {
+		messages: [],
+	};
+
+	for (let i = 0; i < args.length; i++) {
+		const arg = args[i];
+
+		if (arg === "--help" || arg === "-h") {
+			result.help = true;
+		} else if (arg === "--continue" || arg === "-c") {
+			result.continue = true;
+		} else if (arg === "--provider" && i + 1 < args.length) {
+			result.provider = args[++i];
+		} else if (arg === "--model" && i + 1 < args.length) {
+			result.model = args[++i];
+		} else if (arg === "--api-key" && i + 1 < args.length) {
+			result.apiKey = args[++i];
+		} else if (arg === "--system-prompt" && i + 1 < args.length) {
+			result.systemPrompt = args[++i];
+		} else if (!arg.startsWith("-")) {
+			result.messages.push(arg);
+		}
+	}
+
+	return result;
+}
+
+function printHelp() {
+	console.log(`${chalk.bold("coding-agent")} - AI coding assistant with read, bash, edit, write tools
+
+${chalk.bold("Usage:")}
+  coding-agent [options] [messages...]
+
+${chalk.bold("Options:")}
+  --provider <name>       Provider name (default: google)
+  --model <id>            Model ID (default: gemini-2.5-flash)
+  --api-key <key>         API key (defaults to env vars)
+  --system-prompt <text>  System prompt (default: coding assistant prompt)
+  --continue, -c          Continue previous session
+  --help, -h              Show this help
+
+${chalk.bold("Examples:")}
+  # Single message
+  coding-agent "List all .ts files in src/"
+
+  # Multiple messages
+  coding-agent "Read package.json" "What dependencies do we have?"
+
+  # Continue previous session
+  coding-agent --continue "What did we discuss?"
+
+  # Use different model
+  coding-agent --provider openai --model gpt-4o-mini "Help me refactor this code"
+
+${chalk.bold("Environment Variables:")}
+  GEMINI_API_KEY       - Google Gemini API key
+  OPENAI_API_KEY       - OpenAI API key
+  ANTHROPIC_API_KEY    - Anthropic API key
+  CODING_AGENT_DIR     - Session storage directory (default: ~/.coding-agent)
+
+${chalk.bold("Available Tools:")}
+  read   - Read file contents
+  bash   - Execute bash commands
+  edit   - Edit files with find/replace
+  write  - Write files (creates/overwrites)
+`);
+}
+
+const DEFAULT_SYSTEM_PROMPT = `You are an expert coding assistant. You help users with coding tasks by reading files, executing commands, editing code, and writing new files.
+
+Available tools:
+- read: Read file contents
+- bash: Execute bash commands (ls, grep, find, etc.)
+- edit: Make surgical edits to files (find exact text and replace)
+- write: Create or overwrite files
+
+Guidelines:
+- Always use bash tool for file operations like ls, grep, find
+- Use read to examine files before editing
+- Use edit for precise changes (old text must match exactly)
+- Use write only for new files or complete rewrites
+- Be concise in your responses
+- Show file paths clearly when working with files
+
+Current directory: ${process.cwd()}`;
+
+export async function main(args: string[]) {
+	const parsed = parseArgs(args);
+
+	if (parsed.help) {
+		printHelp();
+		return;
+	}
+
+	// Setup session manager
+	const sessionManager = new SessionManager(parsed.continue);
+
+	// Determine provider and model
+	const provider = (parsed.provider || "google") as any;
+	const modelId = parsed.model || "gemini-2.5-flash";
+
+	// Get API key
+	let apiKey = parsed.apiKey;
+	if (!apiKey) {
+		const envVarMap: Record<string, string> = {
+			google: "GEMINI_API_KEY",
+			openai: "OPENAI_API_KEY",
+			anthropic: "ANTHROPIC_API_KEY",
+			xai: "XAI_API_KEY",
+			groq: "GROQ_API_KEY",
+			cerebras: "CEREBRAS_API_KEY",
+			zai: "ZAI_API_KEY",
+		};
+		const envVar = envVarMap[provider] || `${provider.toUpperCase()}_API_KEY`;
+		apiKey = process.env[envVar];
+
+		if (!apiKey) {
+			console.error(chalk.red(`Error: No API key found for provider "${provider}"`));
+			console.error(chalk.dim(`Set ${envVar} environment variable or use --api-key flag`));
+			process.exit(1);
+		}
+	}
+
+	// Create agent
+	const model = getModel(provider, modelId);
+	const systemPrompt = parsed.systemPrompt || DEFAULT_SYSTEM_PROMPT;
+
+	const agent = new Agent({
+		initialState: {
+			systemPrompt,
+			model,
+			thinkingLevel: "off",
+			tools: codingTools,
+		},
+		transport: new ProviderTransport({
+			getApiKey: async () => apiKey!,
+		}),
+	});
+
+	// Load previous messages if continuing
+	if (parsed.continue) {
+		const messages = sessionManager.loadMessages();
+		if (messages.length > 0) {
+			console.log(chalk.dim(`Loaded ${messages.length} messages from previous session`));
+			agent.replaceMessages(messages);
+		}
+	}
+
+	// Start session
+	sessionManager.startSession(agent.state);
+
+	// Subscribe to state updates to save messages
+	agent.subscribe((event) => {
+		if (event.type === "state-update") {
+			// Save any new messages
+			const currentMessages = event.state.messages;
+			const loadedMessages = sessionManager.loadMessages();
+
+			if (currentMessages.length > loadedMessages.length) {
+				for (let i = loadedMessages.length; i < currentMessages.length; i++) {
+					sessionManager.saveMessage(currentMessages[i]);
+				}
+			}
+		}
+
+		sessionManager.saveEvent(event);
+	});
+
+	// Process messages
+	if (parsed.messages.length === 0) {
+		console.log(chalk.yellow("No messages provided. Use --help for usage information."));
+		console.log(chalk.dim(`Session saved to: ${sessionManager.getSessionFile()}`));
+		return;
+	}
+
+	for (const message of parsed.messages) {
+		console.log(chalk.blue(`\n> ${message}\n`));
+		await agent.prompt(message);
+
+		// Print response
+		const lastMessage = agent.state.messages[agent.state.messages.length - 1];
+		if (lastMessage.role === "assistant") {
+			for (const content of lastMessage.content) {
+				if (content.type === "text") {
+					console.log(content.text);
+				}
+			}
+		}
+	}
+
+	console.log(chalk.dim(`\nSession saved to: ${sessionManager.getSessionFile()}`));
+}

packages/coding-agent/src/session-manager.ts

@@ -0,0 +1,167 @@
+import type { AgentEvent, AgentState } from "@mariozechner/pi-agent";
+import { randomBytes } from "crypto";
+import { appendFileSync, existsSync, mkdirSync, readdirSync, readFileSync, statSync } from "fs";
+import { homedir } from "os";
+import { join, resolve } from "path";
+
+function uuidv4(): string {
+	const bytes = randomBytes(16);
+	bytes[6] = (bytes[6] & 0x0f) | 0x40;
+	bytes[8] = (bytes[8] & 0x3f) | 0x80;
+	const hex = bytes.toString("hex");
+	return `${hex.slice(0, 8)}-${hex.slice(8, 12)}-${hex.slice(12, 16)}-${hex.slice(16, 20)}-${hex.slice(20, 32)}`;
+}
+
+export interface SessionHeader {
+	type: "session";
+	id: string;
+	timestamp: string;
+	cwd: string;
+	systemPrompt: string;
+	model: string;
+}
+
+export interface SessionMessageEntry {
+	type: "message";
+	timestamp: string;
+	message: any; // AppMessage from agent state
+}
+
+export interface SessionEventEntry {
+	type: "event";
+	timestamp: string;
+	event: AgentEvent;
+}
+
+export class SessionManager {
+	private sessionId!: string;
+	private sessionFile!: string;
+	private sessionDir: string;
+
+	constructor(continueSession: boolean = false) {
+		this.sessionDir = this.getSessionDirectory();
+
+		if (continueSession) {
+			const mostRecent = this.findMostRecentlyModifiedSession();
+			if (mostRecent) {
+				this.sessionFile = mostRecent;
+				this.loadSessionId();
+			} else {
+				this.initNewSession();
+			}
+		} else {
+			this.initNewSession();
+		}
+	}
+
+	private getSessionDirectory(): string {
+		const cwd = process.cwd();
+		const safePath = "--" + cwd.replace(/^\//, "").replace(/\//g, "-") + "--";
+
+		const configDir = resolve(process.env.CODING_AGENT_DIR || join(homedir(), ".coding-agent"));
+		const sessionDir = join(configDir, "sessions", safePath);
+		if (!existsSync(sessionDir)) {
+			mkdirSync(sessionDir, { recursive: true });
+		}
+		return sessionDir;
+	}
+
+	private initNewSession(): void {
+		this.sessionId = uuidv4();
+		const timestamp = new Date().toISOString().replace(/[:.]/g, "-");
+		this.sessionFile = join(this.sessionDir, `${timestamp}_${this.sessionId}.jsonl`);
+	}
+
+	private findMostRecentlyModifiedSession(): string | null {
+		try {
+			const files = readdirSync(this.sessionDir)
+				.filter((f) => f.endsWith(".jsonl"))
+				.map((f) => ({
+					name: f,
+					path: join(this.sessionDir, f),
+					mtime: statSync(join(this.sessionDir, f)).mtime,
+				}))
+				.sort((a, b) => b.mtime.getTime() - a.mtime.getTime());
+
+			return files[0]?.path || null;
+		} catch {
+			return null;
+		}
+	}
+
+	private loadSessionId(): void {
+		if (!existsSync(this.sessionFile)) return;
+
+		const lines = readFileSync(this.sessionFile, "utf8").trim().split("\n");
+		for (const line of lines) {
+			try {
+				const entry = JSON.parse(line);
+				if (entry.type === "session") {
+					this.sessionId = entry.id;
+					return;
+				}
+			} catch {
+				// Skip malformed lines
+			}
+		}
+		this.sessionId = uuidv4();
+	}
+
+	startSession(state: AgentState): void {
+		const entry: SessionHeader = {
+			type: "session",
+			id: this.sessionId,
+			timestamp: new Date().toISOString(),
+			cwd: process.cwd(),
+			systemPrompt: state.systemPrompt,
+			model: `${state.model.provider}/${state.model.id}`,
+		};
+		appendFileSync(this.sessionFile, JSON.stringify(entry) + "\n");
+	}
+
+	saveMessage(message: any): void {
+		const entry: SessionMessageEntry = {
+			type: "message",
+			timestamp: new Date().toISOString(),
+			message,
+		};
+		appendFileSync(this.sessionFile, JSON.stringify(entry) + "\n");
+	}
+
+	saveEvent(event: AgentEvent): void {
+		const entry: SessionEventEntry = {
+			type: "event",
+			timestamp: new Date().toISOString(),
+			event,
+		};
+		appendFileSync(this.sessionFile, JSON.stringify(entry) + "\n");
+	}
+
+	loadMessages(): any[] {
+		if (!existsSync(this.sessionFile)) return [];
+
+		const messages: any[] = [];
+		const lines = readFileSync(this.sessionFile, "utf8").trim().split("\n");
+
+		for (const line of lines) {
+			try {
+				const entry = JSON.parse(line);
+				if (entry.type === "message") {
+					messages.push(entry.message);
+				}
+			} catch {
+				// Skip malformed lines
+			}
+		}
+
+		return messages;
+	}
+
+	getSessionId(): string {
+		return this.sessionId;
+	}
+
+	getSessionFile(): string {
+		return this.sessionFile;
+	}
+}

packages/coding-agent/src/tools/bash.ts

@@ -0,0 +1,37 @@
+import type { AgentTool } from "@mariozechner/pi-ai";
+import { Type } from "@sinclair/typebox";
+import { exec } from "child_process";
+import { promisify } from "util";
+
+const execAsync = promisify(exec);
+
+const bashSchema = Type.Object({
+	command: Type.String({ description: "Bash command to execute" }),
+});
+
+export const bashTool: AgentTool<typeof bashSchema> = {
+	name: "bash",
+	label: "bash",
+	description:
+		"Execute a bash command in the current working directory. Returns stdout and stderr. Commands run with a 30 second timeout.",
+	parameters: bashSchema,
+	execute: async (_toolCallId: string, { command }: { command: string }) => {
+		try {
+			const { stdout, stderr } = await execAsync(command, {
+				timeout: 30000,
+				maxBuffer: 10 * 1024 * 1024, // 10MB
+			});
+
+			let output = "";
+			if (stdout) output += stdout;
+			if (stderr) output += stderr ? `\nSTDERR:\n${stderr}` : "";
+
+			return { output: output || "(no output)", details: undefined };
+		} catch (error: any) {
+			return {
+				output: `Error executing command: ${error.message}\nSTDOUT: ${error.stdout || ""}\nSTDERR: ${error.stderr || ""}`,
+				details: undefined,
+			};
+		}
+	},
+};

packages/coding-agent/src/tools/edit.ts

@@ -0,0 +1,61 @@
+import type { AgentTool } from "@mariozechner/pi-ai";
+import { Type } from "@sinclair/typebox";
+import { existsSync, readFileSync, writeFileSync } from "fs";
+import { resolve } from "path";
+
+const editSchema = Type.Object({
+	path: Type.String({ description: "Path to the file to edit (relative or absolute)" }),
+	oldText: Type.String({ description: "Exact text to find and replace (must match exactly)" }),
+	newText: Type.String({ description: "New text to replace the old text with" }),
+});
+
+export const editTool: AgentTool<typeof editSchema> = {
+	name: "edit",
+	label: "edit",
+	description:
+		"Edit a file by replacing exact text. The oldText must match exactly (including whitespace). Use this for precise, surgical edits.",
+	parameters: editSchema,
+	execute: async (
+		_toolCallId: string,
+		{ path, oldText, newText }: { path: string; oldText: string; newText: string },
+	) => {
+		try {
+			const absolutePath = resolve(path);
+
+			if (!existsSync(absolutePath)) {
+				return { output: `Error: File not found: ${path}`, details: undefined };
+			}
+
+			const content = readFileSync(absolutePath, "utf-8");
+
+			// Check if old text exists
+			if (!content.includes(oldText)) {
+				return {
+					output: `Error: Could not find the exact text in ${path}. The old text must match exactly including all whitespace and newlines.`,
+					details: undefined,
+				};
+			}
+
+			// Count occurrences
+			const occurrences = content.split(oldText).length - 1;
+
+			if (occurrences > 1) {
+				return {
+					output: `Error: Found ${occurrences} occurrences of the text in ${path}. The text must be unique. Please provide more context to make it unique.`,
+					details: undefined,
+				};
+			}
+
+			// Perform replacement
+			const newContent = content.replace(oldText, newText);
+			writeFileSync(absolutePath, newContent, "utf-8");
+
+			return {
+				output: `Successfully replaced text in ${path}. Changed ${oldText.length} characters to ${newText.length} characters.`,
+				details: undefined,
+			};
+		} catch (error: any) {
+			return { output: `Error editing file: ${error.message}`, details: undefined };
+		}
+	},
+};

packages/coding-agent/src/tools/index.ts

@@ -0,0 +1,11 @@
+export { bashTool } from "./bash.js";
+export { editTool } from "./edit.js";
+export { readTool } from "./read.js";
+export { writeTool } from "./write.js";
+
+import { bashTool } from "./bash.js";
+import { editTool } from "./edit.js";
+import { readTool } from "./read.js";
+import { writeTool } from "./write.js";
+
+export const codingTools = [readTool, bashTool, editTool, writeTool];

packages/coding-agent/src/tools/read.ts

@@ -0,0 +1,29 @@
+import type { AgentTool } from "@mariozechner/pi-ai";
+import { Type } from "@sinclair/typebox";
+import { existsSync, readFileSync } from "fs";
+import { resolve } from "path";
+
+const readSchema = Type.Object({
+	path: Type.String({ description: "Path to the file to read (relative or absolute)" }),
+});
+
+export const readTool: AgentTool<typeof readSchema> = {
+	name: "read",
+	label: "read",
+	description: "Read the contents of a file. Returns the full file content as text.",
+	parameters: readSchema,
+	execute: async (_toolCallId: string, { path }: { path: string }) => {
+		try {
+			const absolutePath = resolve(path);
+
+			if (!existsSync(absolutePath)) {
+				return { output: `Error: File not found: ${path}`, details: undefined };
+			}
+
+			const content = readFileSync(absolutePath, "utf-8");
+			return { output: content, details: undefined };
+		} catch (error: any) {
+			return { output: `Error reading file: ${error.message}`, details: undefined };
+		}
+	},
+};

packages/coding-agent/src/tools/write.ts

@@ -0,0 +1,31 @@
+import type { AgentTool } from "@mariozechner/pi-ai";
+import { Type } from "@sinclair/typebox";
+import { mkdirSync, writeFileSync } from "fs";
+import { dirname, resolve } from "path";
+
+const writeSchema = Type.Object({
+	path: Type.String({ description: "Path to the file to write (relative or absolute)" }),
+	content: Type.String({ description: "Content to write to the file" }),
+});
+
+export const writeTool: AgentTool<typeof writeSchema> = {
+	name: "write",
+	label: "write",
+	description:
+		"Write content to a file. Creates the file if it doesn't exist, overwrites if it does. Automatically creates parent directories.",
+	parameters: writeSchema,
+	execute: async (_toolCallId: string, { path, content }: { path: string; content: string }) => {
+		try {
+			const absolutePath = resolve(path);
+			const dir = dirname(absolutePath);
+
+			// Create parent directories if needed
+			mkdirSync(dir, { recursive: true });
+
+			writeFileSync(absolutePath, content, "utf-8");
+			return { output: `Successfully wrote ${content.length} bytes to ${path}`, details: undefined };
+		} catch (error: any) {
+			return { output: `Error writing file: ${error.message}`, details: undefined };
+		}
+	},
+};

packages/coding-agent/test/tools.test.ts

@@ -0,0 +1,133 @@
+import { mkdirSync, rmSync, writeFileSync } from "fs";
+import { tmpdir } from "os";
+import { join } from "path";
+import { afterEach, beforeEach, describe, expect, it } from "vitest";
+import { bashTool } from "../src/tools/bash.js";
+import { editTool } from "../src/tools/edit.js";
+import { readTool } from "../src/tools/read.js";
+import { writeTool } from "../src/tools/write.js";
+
+describe("Coding Agent Tools", () => {
+	let testDir: string;
+
+	beforeEach(() => {
+		// Create a unique temporary directory for each test
+		testDir = join(tmpdir(), `coding-agent-test-${Date.now()}`);
+		mkdirSync(testDir, { recursive: true });
+	});
+
+	afterEach(() => {
+		// Clean up test directory
+		rmSync(testDir, { recursive: true, force: true });
+	});
+
+	describe("read tool", () => {
+		it("should read file contents", async () => {
+			const testFile = join(testDir, "test.txt");
+			const content = "Hello, world!";
+			writeFileSync(testFile, content);
+
+			const result = await readTool.execute("test-call-1", { path: testFile });
+
+			expect(result.output).toBe(content);
+			expect(result.details).toBeUndefined();
+		});
+
+		it("should handle non-existent files", async () => {
+			const testFile = join(testDir, "nonexistent.txt");
+
+			const result = await readTool.execute("test-call-2", { path: testFile });
+
+			expect(result.output).toContain("Error");
+			expect(result.output).toContain("File not found");
+		});
+	});
+
+	describe("write tool", () => {
+		it("should write file contents", async () => {
+			const testFile = join(testDir, "write-test.txt");
+			const content = "Test content";
+
+			const result = await writeTool.execute("test-call-3", { path: testFile, content });
+
+			expect(result.output).toContain("Successfully wrote");
+			expect(result.output).toContain(testFile);
+			expect(result.details).toBeUndefined();
+		});
+
+		it("should create parent directories", async () => {
+			const testFile = join(testDir, "nested", "dir", "test.txt");
+			const content = "Nested content";
+
+			const result = await writeTool.execute("test-call-4", { path: testFile, content });
+
+			expect(result.output).toContain("Successfully wrote");
+		});
+	});
+
+	describe("edit tool", () => {
+		it("should replace text in file", async () => {
+			const testFile = join(testDir, "edit-test.txt");
+			const originalContent = "Hello, world!";
+			writeFileSync(testFile, originalContent);
+
+			const result = await editTool.execute("test-call-5", {
+				path: testFile,
+				oldText: "world",
+				newText: "testing",
+			});
+
+			expect(result.output).toContain("Successfully replaced");
+			expect(result.details).toBeUndefined();
+		});
+
+		it("should fail if text not found", async () => {
+			const testFile = join(testDir, "edit-test.txt");
+			const originalContent = "Hello, world!";
+			writeFileSync(testFile, originalContent);
+
+			const result = await editTool.execute("test-call-6", {
+				path: testFile,
+				oldText: "nonexistent",
+				newText: "testing",
+			});
+
+			expect(result.output).toContain("Could not find the exact text");
+		});
+
+		it("should fail if text appears multiple times", async () => {
+			const testFile = join(testDir, "edit-test.txt");
+			const originalContent = "foo foo foo";
+			writeFileSync(testFile, originalContent);
+
+			const result = await editTool.execute("test-call-7", {
+				path: testFile,
+				oldText: "foo",
+				newText: "bar",
+			});
+
+			expect(result.output).toContain("Found 3 occurrences");
+		});
+	});
+
+	describe("bash tool", () => {
+		it("should execute simple commands", async () => {
+			const result = await bashTool.execute("test-call-8", { command: "echo 'test output'" });
+
+			expect(result.output).toContain("test output");
+			expect(result.details).toBeUndefined();
+		});
+
+		it("should handle command errors", async () => {
+			const result = await bashTool.execute("test-call-9", { command: "exit 1" });
+
+			expect(result.output).toContain("Command failed");
+		});
+
+		it("should respect timeout", async () => {
+			const result = await bashTool.execute("test-call-10", { command: "sleep 35" });
+
+			expect(result.output).toContain("Command failed");
+		}, 35000);
+	});
+});

packages/coding-agent/tsconfig.build.json

@@ -0,0 +1,9 @@
+{
+	"extends": "../../tsconfig.base.json",
+	"compilerOptions": {
+		"outDir": "./dist",
+		"rootDir": "./src"
+	},
+	"include": ["src/**/*.ts"],
+	"exclude": ["node_modules", "dist", "**/*.d.ts", "src/**/*.d.ts"]
+}

packages/coding-agent/vitest.config.ts

@@ -0,0 +1,9 @@
+import { defineConfig } from 'vitest/config';
+
+export default defineConfig({
+  test: {
+    globals: true,
+    environment: 'node',
+    testTimeout: 30000, // 30 seconds for API calls
+  }
+});

packages/pods/package.json

@@ -34,7 +34,7 @@
 		"node": ">=20.0.0"
 	},
 	"dependencies": {
-		"@mariozechner/pi-agent": "^0.5.44",
+		"@mariozechner/pi-agent-old": "^0.5.44",
 		"chalk": "^5.5.0"
 	},
 	"devDependencies": {}

packages/pods/src/commands/prompt.ts

@@ -1,4 +1,4 @@
-import { main as agentMain } from "@mariozechner/pi-agent";
+import { main as agentMain } from "@mariozechner/pi-agent-old";
 import chalk from "chalk";
 import { getActivePod, loadConfig } from "../config.js";
 

packages/web-ui/example/src/main.ts

@@ -9,6 +9,7 @@ import {
 	type AppMessage,
 	AppStorage,
 	ChatPanel,
+	createJavaScriptReplTool,
 	IndexedDBStorageBackend,
 	// PersistentStorageDialog, // TODO: Fix - currently broken
 	ProviderKeysStore,
@@ -197,6 +198,12 @@ Feel free to use these tools when needed to provide accurate and helpful respons
 	await chatPanel.setAgent(agent, {
 		onApiKeyRequired: async (provider: string) => {
 			return await ApiKeyPromptDialog.prompt(provider);
+		},
+		toolsFactory: (agent, agentInterface, artifactsPanel, runtimeProvidersFactory) => {
+			// Create javascript_repl tool with access to attachments + artifacts
+			const replTool = createJavaScriptReplTool();
+			replTool.runtimeProvidersFactory = runtimeProvidersFactory;
+			return [replTool];
 		}
 	});
 };

packages/web-ui/src/ChatPanel.ts

@@ -9,7 +9,6 @@ import { ArtifactsRuntimeProvider } from "./components/sandbox/ArtifactsRuntimeP
 import { AttachmentsRuntimeProvider } from "./components/sandbox/AttachmentsRuntimeProvider.js";
 import type { SandboxRuntimeProvider } from "./components/sandbox/SandboxRuntimeProvider.js";
 import { ArtifactsPanel, ArtifactsToolRenderer } from "./tools/artifacts/index.js";
-import { createJavaScriptReplTool } from "./tools/javascript-repl.js";
 import { registerToolRenderer } from "./tools/renderer-registry.js";
 import type { Attachment } from "./utils/attachment-utils.js";
 import { i18n } from "./utils/i18n.js";
@@ -65,6 +64,7 @@ export class ChatPanel extends LitElement {
 				agent: Agent,
 				agentInterface: AgentInterface,
 				artifactsPanel: ArtifactsPanel,
+				runtimeProvidersFactory: () => SandboxRuntimeProvider[],
 			) => AgentTool<any>[];
 		},
 	) {
@@ -80,12 +80,6 @@ export class ChatPanel extends LitElement {
 		this.agentInterface.onApiKeyRequired = config?.onApiKeyRequired;
 		this.agentInterface.onBeforeSend = config?.onBeforeSend;
 
-		// Create JavaScript REPL tool
-		const javascriptReplTool = createJavaScriptReplTool();
-		if (config?.sandboxUrlProvider) {
-			javascriptReplTool.sandboxUrlProvider = config.sandboxUrlProvider;
-		}
-
 		// Set up artifacts panel
 		this.artifactsPanel = new ArtifactsPanel();
 		if (config?.sandboxUrlProvider) {
@@ -94,7 +88,7 @@ export class ChatPanel extends LitElement {
 		// Register the standalone tool renderer (not the panel itself)
 		registerToolRenderer("artifacts", new ArtifactsToolRenderer(this.artifactsPanel));
 
-		// Runtime providers factory
+		// Runtime providers factory for attachments + artifacts access
 		const runtimeProvidersFactory = () => {
 			const attachments: Attachment[] = [];
 			for (const message of this.agent!.state.messages) {
@@ -116,7 +110,6 @@ export class ChatPanel extends LitElement {
 
 			return providers;
 		};
-		javascriptReplTool.runtimeProvidersFactory = runtimeProvidersFactory;
 		this.artifactsPanel.runtimeProvidersFactory = runtimeProvidersFactory;
 
 		this.artifactsPanel.onArtifactsChange = () => {
@@ -141,8 +134,10 @@ export class ChatPanel extends LitElement {
 		};
 
 		// Set tools on the agent
-		const additionalTools = config?.toolsFactory?.(agent, this.agentInterface, this.artifactsPanel) || [];
-		const tools = [javascriptReplTool, this.artifactsPanel.tool, ...additionalTools];
+		// Pass runtimeProvidersFactory so consumers can configure their own REPL tools
+		const additionalTools =
+			config?.toolsFactory?.(agent, this.agentInterface, this.artifactsPanel, runtimeProvidersFactory) || [];
+		const tools = [this.artifactsPanel.tool, ...additionalTools];
 		this.agent.setTools(tools);
 
 		// Reconstruct artifacts from existing messages

packages/web-ui/src/components/sandbox/ArtifactsRuntimeProvider.ts

@@ -1,4 +1,4 @@
-import { ARTIFACTS_RUNTIME_PROVIDER_DESCRIPTION } from "../../prompts/tool-prompts.js";
+import { ARTIFACTS_RUNTIME_PROVIDER_DESCRIPTION } from "../../prompts/prompts.js";
 import type { SandboxRuntimeProvider } from "./SandboxRuntimeProvider.js";
 
 // Define minimal interface for ArtifactsPanel to avoid circular dependencies

packages/web-ui/src/components/sandbox/AttachmentsRuntimeProvider.ts

@@ -1,4 +1,4 @@
-import { ATTACHMENTS_RUNTIME_DESCRIPTION } from "../../prompts/tool-prompts.js";
+import { ATTACHMENTS_RUNTIME_DESCRIPTION } from "../../prompts/prompts.js";
 import type { Attachment } from "../../utils/attachment-utils.js";
 import type { SandboxRuntimeProvider } from "./SandboxRuntimeProvider.js";
 

packages/web-ui/src/components/sandbox/ConsoleRuntimeProvider.ts

@@ -23,6 +23,10 @@ export class ConsoleRuntimeProvider implements SandboxRuntimeProvider {
 		return {};
 	}
 
+	getDescription(): string {
+		return "";
+	}
+
 	getRuntime(): (sandboxId: string) => void {
 		return (_sandboxId: string) => {
 			// Store truly original console methods on first wrap only

packages/web-ui/src/components/sandbox/SandboxRuntimeProvider.ts

@@ -31,5 +31,5 @@ export interface SandboxRuntimeProvider {
 	 * Optional documentation describing what globals/functions this provider injects.
 	 * This will be appended to tool descriptions dynamically so the LLM knows what's available.
 	 */
-	getDescription?(): string;
+	getDescription(): string;
 }

packages/web-ui/src/index.ts

@@ -56,8 +56,8 @@ export { ApiKeysTab, ProxyTab, SettingsDialog, SettingsTab } from "./dialogs/Set
 // Prompts
 export {
 	ARTIFACTS_RUNTIME_PROVIDER_DESCRIPTION,
-	DOWNLOADABLE_FILE_RUNTIME_DESCRIPTION,
-} from "./prompts/tool-prompts.js";
+	ATTACHMENTS_RUNTIME_DESCRIPTION,
+} from "./prompts/prompts.js";
 // Storage
 export { AppStorage, getAppStorage, setAppStorage } from "./storage/app-storage.js";
 export { IndexedDBStorageBackend } from "./storage/backends/indexeddb-storage-backend.js";

packages/web-ui/src/prompts/prompts.ts

@@ -7,7 +7,7 @@
 // JavaScript REPL Tool
 // ============================================================================
 
-export const JAVASCRIPT_REPL_DESCRIPTION = `# JavaScript REPL
+export const JAVASCRIPT_REPL_TOOL_DESCRIPTION = (runtimeProviderDescriptions: string[]) => `# JavaScript REPL
 
 ## Purpose
 Execute JavaScript code in a sandboxed browser environment with full Web APIs.
@@ -16,7 +16,7 @@ Execute JavaScript code in a sandboxed browser environment with full Web APIs.
 - Quick calculations or data transformations
 - Testing JavaScript code snippets in isolation
 - Processing data with libraries (XLSX, CSV, etc.)
-- Creating visualizations (charts, graphs)
+- Creating artifacts from data
 
 ## Environment
 - ES2023+ JavaScript (async/await, optional chaining, nullish coalescing, etc.)
@@ -54,13 +54,21 @@ console.log('Sum:', sum, 'Average:', avg);
 ## Important Notes
 - Graphics: Use fixed dimensions (800x600), NOT window.innerWidth/Height
 - Chart.js: Set options: { responsive: false, animation: false }
-- Three.js: renderer.setSize(800, 600) with matching aspect ratio`;
+- Three.js: renderer.setSize(800, 600) with matching aspect ratio
+
+## Library functions
+You can use the following functions in your code:
+
+${runtimeProviderDescriptions.join("\n\n")}
+`;
 
 // ============================================================================
 // Artifacts Tool
 // ============================================================================
 
-export const ARTIFACTS_BASE_DESCRIPTION = `Creates and manages file artifacts. Each artifact is a file with a filename and content.
+export const ARTIFACTS_TOOL_DESCRIPTION = (
+	runtimeProviderDescriptions: string[],
+) => `Creates and manages file artifacts. Each artifact is a file with a filename and content.
 
 CRITICAL - ARTIFACT UPDATE WORKFLOW:
 1. Creating new file? → Use 'create'
@@ -104,33 +112,8 @@ Commands:
 ANTI-PATTERNS TO AVOID:
 ❌ Using 'get' + modifying content + 'rewrite' to change one section
 ❌ Using createOrUpdateArtifact() in code for manual edits YOU make
-✅ Use 'update' command for surgical, targeted modifications`;
+✅ Use 'update' command for surgical, targeted modifications
 
-export const ARTIFACTS_RUNTIME_EXAMPLE = `- Example HTML artifact that processes a CSV attachment:
-  <script>
-    // List available files
-    const files = listAttachments();
-    console.log('Available files:', files);
-
-    // Find CSV file
-    const csvFile = files.find(f => f.mimeType === 'text/csv');
-    if (csvFile) {
-      const csvContent = readTextAttachment(csvFile.id);
-      // Process CSV data...
-    }
-
-    // Display image
-    const imageFile = files.find(f => f.mimeType.startsWith('image/'));
-    if (imageFile) {
-      const bytes = readBinaryAttachment(imageFile.id);
-      const blob = new Blob([bytes], {type: imageFile.mimeType});
-      const url = URL.createObjectURL(blob);
-      document.body.innerHTML = '<img src="' + url + '">';
-    }
-  </script>
-`;
-
-export const ARTIFACTS_HTML_SECTION = `
 For text/html artifacts:
 - Must be a single self-contained file
 - External scripts: Use CDNs like https://esm.sh, https://unpkg.com, or https://cdnjs.cloudflare.com
@@ -166,40 +149,33 @@ CRITICAL REMINDER FOR ALL ARTIFACTS:
 - Prefer to update existing files rather than creating new ones
 - Keep filenames consistent and descriptive
 - Use appropriate file extensions
-- Ensure HTML artifacts have a defined background color`;
+- Ensure HTML artifacts have a defined background color
 
-/**
- * Build complete artifacts description with optional provider docs.
- */
-export function buildArtifactsDescription(providerDocs?: string): string {
-	const runtimeSection = providerDocs
-		? `
+The following functions are available inside your code in HTML artifacts:
 
-For text/html artifacts with runtime capabilities:${providerDocs}
-${ARTIFACTS_RUNTIME_EXAMPLE}
-`
-		: "";
-
-	return ARTIFACTS_BASE_DESCRIPTION + runtimeSection + ARTIFACTS_HTML_SECTION;
-}
+${runtimeProviderDescriptions.join("\n\n")}
+`;
 
 // ============================================================================
 // Artifacts Runtime Provider
 // ============================================================================
 
 export const ARTIFACTS_RUNTIME_PROVIDER_DESCRIPTION = `
-Artifact Management from within executed code (HTML/JavaScript REPL).
+### Artifacts
 
-WHEN TO USE THESE FUNCTIONS:
+Programmatically create, read, update, and delete artifact files from your code.
+
+#### When to Use
+- Persist data or state between REPL calls
 - ONLY when writing code that programmatically generates/transforms data
 - Examples: Web scraping results, processed CSV data, generated charts saved as JSON
 - The artifact content is CREATED BY THE CODE, not by you directly
 
-DO NOT USE THESE FUNCTIONS FOR:
+#### Do NOT Use For
 - Summaries or notes YOU write (use artifacts tool instead)
 - Content YOU author directly (use artifacts tool instead)
 
-Functions:
+#### Functions
 - await listArtifacts() - Get list of all artifact filenames, returns string[]
   * Example: const files = await listArtifacts(); // ['data.json', 'notes.md']
 
@@ -216,39 +192,75 @@ Functions:
 - await deleteArtifact(filename) - Delete an artifact
   * Example: await deleteArtifact('temp.json')
 
-Example - Scraping data and saving it:
-  const response = await fetch('https://api.example.com/data');
-  const data = await response.json();
-  await createOrUpdateArtifact('api-results.json', data);
+#### Example
+Scraping data and saving it:
+\`\`\`javascript
+const response = await fetch('https://api.example.com/data');
+const data = await response.json();
+await createOrUpdateArtifact('api-results.json', data);
+\`\`\`
 
-Binary data must be converted to a base64 string before passing to createOrUpdateArtifact.
-Example:
-  const blob = await new Promise(resolve => canvas.toBlob(resolve, 'image/png'));
-  const arrayBuffer = await blob.arrayBuffer();
-  const base64 = btoa(String.fromCharCode(...new Uint8Array(arrayBuffer)));
-  await createOrUpdateArtifact('image.png', base64);
+Binary data (convert to base64 first):
+\`\`\`javascript
+const blob = await new Promise(resolve => canvas.toBlob(resolve, 'image/png'));
+const arrayBuffer = await blob.arrayBuffer();
+const base64 = btoa(String.fromCharCode(...new Uint8Array(arrayBuffer)));
+await createOrUpdateArtifact('image.png', base64);
+\`\`\`
 `;
 
 // ============================================================================
-// Downloadable File Runtime Provider
+// Attachments Runtime Provider
 // ============================================================================
 
-export const DOWNLOADABLE_FILE_RUNTIME_DESCRIPTION = `
-Downloadable Files (one-time downloads for the user - YOU cannot read these back):
-- await returnDownloadableFile(filename, content, mimeType?) - Create downloadable file (async!)
-  * Use for: Processed/transformed data, generated images, analysis results
-  * Important: This creates a download for the user. You will NOT be able to access this file's content later.
-  * If you need to access the data later, use createArtifact() instead (if available).
-  * Always use await with returnDownloadableFile
-  * REQUIRED: For Blob/Uint8Array binary content, you MUST supply a proper MIME type (e.g., "image/png").
-    If omitted, throws an Error with stack trace pointing to the offending line.
-  * Strings without a MIME default to text/plain.
-  * Objects are auto-JSON stringified and default to application/json unless a MIME is provided.
-  * Canvas images: Use toBlob() with await Promise wrapper
-  * Examples:
-    - await returnDownloadableFile('cleaned-data.csv', csvString, 'text/csv')
-    - await returnDownloadableFile('analysis.json', {results: [...]}, 'application/json')
-    - await returnDownloadableFile('chart.png', blob, 'image/png')`;
+export const ATTACHMENTS_RUNTIME_DESCRIPTION = `
+### User Attachments
+
+Read files that the user has uploaded to the conversation.
+
+#### When to Use
+- When you need to read or process files the user has uploaded to the conversation
+- Examples: CSV data files, JSON datasets, Excel spreadsheets, images, PDFs
+
+#### Do NOT Use For
+- Creating new files (use createOrUpdateArtifact instead)
+- Modifying existing files (read first, then create artifact with modified version)
+
+#### Functions
+- listAttachments() - List all attachments, returns array of {id, fileName, mimeType, size}
+  * Example: const files = listAttachments(); // [{id: '...', fileName: 'data.xlsx', mimeType: '...', size: 12345}]
+
+- readTextAttachment(attachmentId) - Read attachment as text, returns string
+  * Use for: CSV, JSON, TXT, XML, and other text-based files
+  * Example: const csvContent = readTextAttachment(files[0].id);
+  * Example: const json = JSON.parse(readTextAttachment(jsonFile.id));
+
+- readBinaryAttachment(attachmentId) - Read attachment as binary data, returns Uint8Array
+  * Use for: Excel (.xlsx), images, PDFs, and other binary files
+  * Example: const xlsxBytes = readBinaryAttachment(files[0].id);
+  * Example: const XLSX = await import('https://esm.run/xlsx'); const workbook = XLSX.read(xlsxBytes);
+
+#### Example
+Processing CSV attachment:
+\`\`\`javascript
+const files = listAttachments();
+const csvFile = files.find(f => f.fileName.endsWith('.csv'));
+const csvData = readTextAttachment(csvFile.id);
+const rows = csvData.split('\\n').map(row => row.split(','));
+console.log(\`Found \${rows.length} rows\`);
+\`\`\`
+
+Processing Excel attachment:
+\`\`\`javascript
+const XLSX = await import('https://esm.run/xlsx');
+const files = listAttachments();
+const excelFile = files.find(f => f.fileName.endsWith('.xlsx'));
+const bytes = readBinaryAttachment(excelFile.id);
+const workbook = XLSX.read(bytes);
+const firstSheet = workbook.Sheets[workbook.SheetNames[0]];
+const jsonData = XLSX.utils.sheet_to_json(firstSheet);
+\`\`\`
+`;
 
 // ============================================================================
 // Extract Document Tool
@@ -273,27 +285,3 @@ Structured plain text with page/sheet/slide delimiters in XML-like format:
 - Maximum file size: 50MB
 - CORS restrictions may block some URLs - if this happens, the error will guide you to help the user configure a CORS proxy
 - Format is automatically detected from file extension and Content-Type header`;
-
-// ============================================================================
-// Attachments Runtime Provider
-// ============================================================================
-
-export const ATTACHMENTS_RUNTIME_DESCRIPTION = `
-User Attachments (files the user added to the conversation):
-- listAttachments() - List all attachments, returns array of {id, fileName, mimeType, size}
-  * Example: const files = listAttachments(); // [{id: '...', fileName: 'data.xlsx', mimeType: '...', size: 12345}]
-- readTextAttachment(attachmentId) - Read attachment as text, returns string
-  * Use for: CSV, JSON, TXT, XML, and other text-based files
-  * Example: const csvContent = readTextAttachment(files[0].id);
-  * Example: const json = JSON.parse(readTextAttachment(jsonFile.id));
-- readBinaryAttachment(attachmentId) - Read attachment as binary data, returns Uint8Array
-  * Use for: Excel (.xlsx), images, PDFs, and other binary files
-  * Example: const xlsxBytes = readBinaryAttachment(files[0].id);
-  * Example: const XLSX = await import('https://esm.run/xlsx'); const workbook = XLSX.read(xlsxBytes);
-
-Common pattern - Process attachment and create download:
-  const files = listAttachments();
-  const csvFile = files.find(f => f.fileName.endsWith('.csv'));
-  const csvData = readTextAttachment(csvFile.id);
-  // Process csvData...
-  await returnDownloadableFile('processed-' + csvFile.fileName, processedData, 'text/csv');`;

packages/web-ui/src/tools/artifacts/artifacts.ts

@@ -8,7 +8,7 @@ import { createRef, type Ref, ref } from "lit/directives/ref.js";
 import { X } from "lucide";
 import type { ArtifactMessage } from "../../components/Messages.js";
 import type { SandboxRuntimeProvider } from "../../components/sandbox/SandboxRuntimeProvider.js";
-import { buildArtifactsDescription } from "../../prompts/tool-prompts.js";
+import { ARTIFACTS_TOOL_DESCRIPTION } from "../../prompts/prompts.js";
 import { i18n } from "../../utils/i18n.js";
 import type { ArtifactElement } from "./ArtifactElement.js";
 import { DocxArtifact } from "./DocxArtifact.js";
@@ -245,14 +245,12 @@ export class ArtifactsPanel extends LitElement {
 			label: "Artifacts",
 			name: "artifacts",
 			get description() {
-				// Get dynamic provider descriptions
-				const providers = self.runtimeProvidersFactory?.() || [];
-				const providerDocs = providers
-					.map((p) => p.getDescription?.())
-					.filter(Boolean)
-					.join("\n");
-
-				return buildArtifactsDescription(providerDocs || undefined);
+				const runtimeProviderDescriptions =
+					self
+						.runtimeProvidersFactory?.()
+						.map((d) => d.getDescription())
+						.filter((d) => d.trim().length > 0) || [];
+				return ARTIFACTS_TOOL_DESCRIPTION(runtimeProviderDescriptions);
 			},
 			parameters: artifactsParamsSchema,
 			// Execute mutates our local store and returns a plain output

packages/web-ui/src/tools/extract-document.ts

@@ -3,7 +3,7 @@ import type { AgentTool, ToolResultMessage } from "@mariozechner/pi-ai";
 import { type Static, Type } from "@sinclair/typebox";
 import { createRef, ref } from "lit/directives/ref.js";
 import { FileText } from "lucide";
-import { EXTRACT_DOCUMENT_DESCRIPTION } from "../prompts/tool-prompts.js";
+import { EXTRACT_DOCUMENT_DESCRIPTION } from "../prompts/prompts.js";
 import { loadAttachment } from "../utils/attachment-utils.js";
 import { registerToolRenderer, renderCollapsibleHeader, renderHeader } from "./renderer-registry.js";
 import type { ToolRenderer, ToolRenderResult } from "./types.js";

packages/web-ui/src/tools/javascript-repl.ts

@@ -5,7 +5,7 @@ import { createRef, ref } from "lit/directives/ref.js";
 import { Code } from "lucide";
 import { type SandboxFile, SandboxIframe, type SandboxResult } from "../components/SandboxedIframe.js";
 import type { SandboxRuntimeProvider } from "../components/sandbox/SandboxRuntimeProvider.js";
-import { JAVASCRIPT_REPL_DESCRIPTION } from "../prompts/tool-prompts.js";
+import { JAVASCRIPT_REPL_TOOL_DESCRIPTION } from "../prompts/prompts.js";
 import type { Attachment } from "../utils/attachment-utils.js";
 import { registerToolRenderer, renderCollapsibleHeader, renderHeader } from "./renderer-registry.js";
 import type { ToolRenderer, ToolRenderResult } from "./types.js";
@@ -132,7 +132,13 @@ export function createJavaScriptReplTool(): AgentTool<typeof javascriptReplSchem
 		name: "javascript_repl",
 		runtimeProvidersFactory: () => [], // default to empty array
 		sandboxUrlProvider: undefined, // optional, for browser extensions
-		description: JAVASCRIPT_REPL_DESCRIPTION,
+		get description() {
+			const runtimeProviderDescriptions =
+				this.runtimeProvidersFactory?.()
+					.map((d) => d.getDescription())
+					.filter((d) => d.trim().length > 0) || [];
+			return JAVASCRIPT_REPL_TOOL_DESCRIPTION(runtimeProviderDescriptions);
+		},
 		parameters: javascriptReplSchema,
 		execute: async function (_toolCallId: string, args: Static<typeof javascriptReplSchema>, signal?: AbortSignal) {
 			const result = await executeJavaScript(

tsconfig.json

@@ -7,6 +7,8 @@
 			"@mariozechner/pi-ai": ["./packages/ai/src/index.ts"],
 			"@mariozechner/pi-web-ui": ["./packages/web-ui/src/index.ts"],
 			"@mariozechner/pi-agent": ["./packages/agent/src/index.ts"],
+			"@mariozechner/pi-agent-old": ["./packages/agent-old/src/index.ts"],
+			"@mariozechner/coding-agent": ["./packages/coding-agent/src/index.ts"],
 			"@mariozechner/pi": ["./packages/pods/src/index.ts"]
 		}
 	},