getcompanion-ai/co-mono

Fork 0

mirror of https://github.com/getcompanion-ai/co-mono.git synced 2026-04-15 18:01:22 +00:00

Mario Zechner e9f6de7cb8 WP14: Create main-new.ts and cli-new.ts using AgentSession and new modes

2025-12-09 00:21:18 +01:00

49 KiB

Raw Blame History

Coding Agent Refactoring Plan

Status

Branch: refactor Started: 2024-12-08

To resume work on this refactoring:

Read this document fully
Run git diff to see current work in progress
Check the work packages below - find first unchecked item
Read any files mentioned in that work package before making changes

Strategy: Keep Old Code for Reference

We create new files alongside old ones instead of modifying in place:

src/modes/print-mode.ts (new) - old code stays in main.ts
src/modes/rpc-mode.ts (new) - old code stays in main.ts
src/modes/interactive/interactive-mode.ts (new) - old code stays in tui/tui-renderer.ts
src/main-new.ts (new) - old code stays in main.ts
src/cli-new.ts (new) - old code stays in cli.ts

This allows:

Parallel comparison of old vs new behavior
Gradual migration and testing
Easy rollback if needed

Final switchover: When everything works, rename files and delete old code.

Goals

Eliminate code duplication between the three run modes (interactive, print/json, rpc)
Create a testable core (AgentSession) that encapsulates all agent/session logic
Separate concerns: TUI rendering vs agent state management vs I/O
Improve naming: TuiRenderer → InteractiveMode (it's not just a renderer)
Simplify main.ts: Move setup logic out, make it just arg parsing + mode routing

Architecture Overview

Current State (Problems)

main.ts (1100+ lines)
├── parseArgs, printHelp
├── buildSystemPrompt, loadProjectContextFiles
├── resolveModelScope, model resolution logic
├── runInteractiveMode() - thin wrapper around TuiRenderer
├── runSingleShotMode() - duplicates event handling, session saving
├── runRpcMode() - duplicates event handling, session saving, auto-compaction, bash execution
└── executeRpcBashCommand() - duplicate of TuiRenderer.executeBashCommand()

tui/tui-renderer.ts (2400+ lines)
├── TUI lifecycle (init, render, event loop)
├── Agent event handling + session persistence (duplicated in main.ts)
├── Auto-compaction logic (duplicated in main.ts runRpcMode)
├── Bash execution (duplicated in main.ts)
├── All slash command implementations (/export, /copy, /model, /thinking, etc.)
├── All hotkey handlers (Ctrl+C, Ctrl+P, Shift+Tab, etc.)
├── Model/thinking cycling logic
└── 6 different selector UIs (model, thinking, theme, session, branch, oauth)

Target State

src/
├── main.ts (~200 lines)
│   ├── parseArgs, printHelp
│   └── Route to appropriate mode
│
├── core/
│   ├── agent-session.ts      # Shared agent/session logic (THE key abstraction)
│   ├── bash-executor.ts      # Bash execution with streaming + cancellation
│   └── setup.ts              # Model resolution, system prompt building, session loading
│
└── modes/
    ├── print-mode.ts         # Simple: prompt, output result
    ├── rpc-mode.ts           # JSON stdin/stdout protocol
    └── interactive/
        ├── interactive-mode.ts   # Main orchestrator
        ├── command-handlers.ts   # Slash command implementations
        ├── hotkeys.ts            # Hotkey handling
        └── selectors.ts          # Modal selector management

AgentSession API

This is the core abstraction shared by all modes. See full API design below.

class AgentSession {
  // ─── Read-only State Access ───
  get state(): AgentState;
  get model(): Model<any> | null;
  get thinkingLevel(): ThinkingLevel;
  get isStreaming(): boolean;
  get messages(): AppMessage[];  // Includes custom types like BashExecutionMessage
  get queueMode(): QueueMode;

  // ─── Event Subscription ───
  // Handles session persistence internally (saves messages, checks auto-compaction)
  subscribe(listener: (event: AgentEvent) => void): () => void;

  // ─── Prompting ───
  prompt(text: string, options?: PromptOptions): Promise<void>;
  queueMessage(text: string): Promise<void>;
  clearQueue(): string[];
  abort(): Promise<void>;
  reset(): Promise<void>;

  // ─── Model Management ───
  setModel(model: Model<any>): Promise<void>;  // Validates API key, saves to session + settings
  cycleModel(): Promise<ModelCycleResult | null>;
  getAvailableModels(): Promise<Model<any>[]>;

  // ─── Thinking Level ───
  setThinkingLevel(level: ThinkingLevel): void;  // Saves to session + settings
  cycleThinkingLevel(): ThinkingLevel | null;
  supportsThinking(): boolean;

  // ─── Queue Mode ───
  setQueueMode(mode: QueueMode): void;  // Saves to settings

  // ─── Compaction ───
  compact(customInstructions?: string): Promise<CompactionResult>;
  abortCompaction(): void;
  checkAutoCompaction(): Promise<CompactionResult | null>;  // Called internally after assistant messages
  setAutoCompactionEnabled(enabled: boolean): void;  // Saves to settings
  get autoCompactionEnabled(): boolean;

  // ─── Bash Execution ───
  executeBash(command: string, onChunk?: (chunk: string) => void): Promise<BashResult>;
  abortBash(): void;
  get isBashRunning(): boolean;

  // Session management
  switchSession(sessionPath: string): Promise<void>;
  branch(entryIndex: number): string;
  getUserMessagesForBranching(): Array<{ entryIndex: number; text: string }>;
  getSessionStats(): SessionStats;
  exportToHtml(outputPath?: string): string;

  // Utilities
  getLastAssistantText(): string | null;
}

Work Packages

WP1: Create bash-executor.ts

Extract bash execution into a standalone module that both AgentSession and tests can use.

Files to create:

src/core/bash-executor.ts

Extract from:

src/tui/tui-renderer.ts: executeBashCommand() method (lines ~2190-2270)
src/main.ts: executeRpcBashCommand() function (lines ~640-700)

Implementation:

// src/core/bash-executor.ts
export interface BashExecutorOptions {
  onChunk?: (chunk: string) => void;
  signal?: AbortSignal;
}

export interface BashResult {
  output: string;
  exitCode: number | null;
  cancelled: boolean;
  truncated: boolean;
  fullOutputPath?: string;
}

export function executeBash(command: string, options?: BashExecutorOptions): Promise<BashResult>;

Logic to include:

Spawn shell process with getShellConfig()
Stream stdout/stderr through onChunk callback (if provided)
Handle temp file creation for large output (> DEFAULT_MAX_BYTES)
Sanitize output (stripAnsi, sanitizeBinaryOutput, normalize newlines)
Apply truncation via truncateTail()
Support cancellation via AbortSignal (calls killProcessTree)
Return structured result

Verification:

npm run check passes
Manual test: Run pi in interactive mode, execute !ls -la, verify output appears
Manual test: Run !sleep 10, press Esc, verify cancellation works

Create src/core/bash-executor.ts with executeBash() function
Add proper TypeScript types and exports
Verify with npm run check

WP2: Create agent-session.ts (Core Structure)

Create the AgentSession class with basic structure and state access.

Files to create:

src/core/agent-session.ts
src/core/index.ts (barrel export)

Dependencies: None (can use existing imports)

Implementation - Phase 1 (structure + state access):

// src/core/agent-session.ts
import type { Agent, AgentEvent, AgentState, AppMessage, ThinkingLevel } from "@mariozechner/pi-agent-core";
import type { Model } from "@mariozechner/pi-ai";
import type { SessionManager } from "../session-manager.js";
import type { SettingsManager } from "../settings-manager.js";

export interface AgentSessionConfig {
  agent: Agent;
  sessionManager: SessionManager;
  settingsManager: SettingsManager;
  scopedModels?: Array<{ model: Model<any>; thinkingLevel: ThinkingLevel }>;
  fileCommands?: FileSlashCommand[];
}

export class AgentSession {
  readonly agent: Agent;
  readonly sessionManager: SessionManager;
  readonly settingsManager: SettingsManager;
  
  private scopedModels: Array<{ model: Model<any>; thinkingLevel: ThinkingLevel }>;
  private fileCommands: FileSlashCommand[];

  constructor(config: AgentSessionConfig) {
    this.agent = config.agent;
    this.sessionManager = config.sessionManager;
    this.settingsManager = config.settingsManager;
    this.scopedModels = config.scopedModels ?? [];
    this.fileCommands = config.fileCommands ?? [];
  }

  // State access (simple getters)
  get state(): AgentState { return this.agent.state; }
  get model(): Model<any> | null { return this.agent.state.model; }
  get thinkingLevel(): ThinkingLevel { return this.agent.state.thinkingLevel; }
  get isStreaming(): boolean { return this.agent.state.isStreaming; }
  get messages(): AppMessage[] { return this.agent.state.messages; }
  get sessionFile(): string { return this.sessionManager.getSessionFile(); }
  get sessionId(): string { return this.sessionManager.getSessionId(); }
}

Verification:

npm run check passes
Class can be instantiated (will test via later integration)

Create src/core/agent-session.ts with basic structure
Create src/core/index.ts barrel export
Verify with npm run check

WP3: AgentSession - Event Subscription + Session Persistence

Add subscribe() method that wraps agent subscription and handles session persistence.

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: subscribeToAgent() method (lines ~470-495)
src/main.ts: runRpcMode() subscription logic (lines ~720-745)
src/main.ts: runSingleShotMode() subscription logic (lines ~605-610)

Implementation:

// Add to AgentSession class

private unsubscribeAgent?: () => void;
private eventListeners: Array<(event: AgentEvent) => void> = [];

/**
 * Subscribe to agent events. Session persistence is handled internally.
 * Multiple listeners can be added. Returns unsubscribe function.
 */
subscribe(listener: (event: AgentEvent) => void): () => void {
  this.eventListeners.push(listener);
  
  // Set up agent subscription if not already done
  if (!this.unsubscribeAgent) {
    this.unsubscribeAgent = this.agent.subscribe(async (event) => {
      // Notify all listeners
      for (const l of this.eventListeners) {
        l(event);
      }
      
      // Handle session persistence
      if (event.type === "message_end") {
        this.sessionManager.saveMessage(event.message);
        
        // Initialize session after first user+assistant exchange
        if (this.sessionManager.shouldInitializeSession(this.agent.state.messages)) {
          this.sessionManager.startSession(this.agent.state);
        }
        
        // Check auto-compaction after assistant messages
        if (event.message.role === "assistant") {
          await this.checkAutoCompaction();
        }
      }
    });
  }
  
  // Return unsubscribe function for this specific listener
  return () => {
    const index = this.eventListeners.indexOf(listener);
    if (index !== -1) {
      this.eventListeners.splice(index, 1);
    }
  };
}

/**
 * Unsubscribe from agent entirely (used during cleanup/reset)
 */
private unsubscribeAll(): void {
  if (this.unsubscribeAgent) {
    this.unsubscribeAgent();
    this.unsubscribeAgent = undefined;
  }
  this.eventListeners = [];
}

Verification:

npm run check passes

Add subscribe() method to AgentSession
Add _disconnectFromAgent() private method (renamed from unsubscribeAll)
Add _reconnectToAgent() private method (renamed from resubscribe)
Add dispose() public method for full cleanup
Verify with npm run check

WP4: AgentSession - Prompting Methods

Add prompt(), queueMessage(), clearQueue(), abort(), reset() methods.

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: editor.onSubmit validation logic (lines ~340-380)
src/tui/tui-renderer.ts: handleClearCommand() (lines ~2005-2035)
Slash command expansion from expandSlashCommand()

Implementation:

// Add to AgentSession class

private queuedMessages: string[] = [];

/**
 * Send a prompt to the agent.
 * - Validates model and API key
 * - Expands slash commands by default
 * - Throws if no model or no API key
 */
async prompt(text: string, options?: { 
  expandSlashCommands?: boolean; 
  attachments?: Attachment[];
}): Promise<void> {
  const expandCommands = options?.expandSlashCommands ?? true;
  
  // Validate model
  if (!this.model) {
    throw new Error(
      "No model selected.\n\n" +
      "Set an API key (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.)\n" +
      `or create ${getModelsPath()}\n\n` +
      "Then use /model to select a model."
    );
  }
  
  // Validate API key
  const apiKey = await getApiKeyForModel(this.model);
  if (!apiKey) {
    throw new Error(
      `No API key found for ${this.model.provider}.\n\n` +
      `Set the appropriate environment variable or update ${getModelsPath()}`
    );
  }
  
  // Expand slash commands
  const expandedText = expandCommands ? expandSlashCommand(text, this.fileCommands) : text;
  
  await this.agent.prompt(expandedText, options?.attachments);
}

/**
 * Queue a message while agent is streaming.
 */
async queueMessage(text: string): Promise<void> {
  this.queuedMessages.push(text);
  await this.agent.queueMessage({
    role: "user",
    content: [{ type: "text", text }],
    timestamp: Date.now(),
  });
}

/**
 * Clear queued messages. Returns them for restoration to editor.
 */
clearQueue(): string[] {
  const queued = [...this.queuedMessages];
  this.queuedMessages = [];
  this.agent.clearMessageQueue();
  return queued;
}

/**
 * Abort current operation and wait for idle.
 */
async abort(): Promise<void> {
  this.agent.abort();
  await this.agent.waitForIdle();
}

/**
 * Reset agent and session. Starts a fresh session.
 */
async reset(): Promise<void> {
  this.unsubscribeAll();
  await this.abort();
  this.agent.reset();
  this.sessionManager.reset();
  this.queuedMessages = [];
  // Re-subscribe (caller may have added listeners before reset)
  // Actually, listeners are cleared in unsubscribeAll, so caller needs to re-subscribe
}

Verification:

npm run check passes

Add prompt() method with validation and slash command expansion
Add queueMessage() method
Add clearQueue() method
Add abort() method
Add reset() method
Add queuedMessageCount getter and getQueuedMessages() method
Verify with npm run check

WP5: AgentSession - Model Management

Add setModel(), cycleModel(), getAvailableModels() methods.

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: cycleModel() method (lines ~970-1070)
Model validation scattered throughout

Implementation:

// Add to AgentSession class

export interface ModelCycleResult {
  model: Model<any>;
  thinkingLevel: ThinkingLevel;
  isScoped: boolean;
}

/**
 * Set model directly. Validates API key, saves to session and settings.
 */
async setModel(model: Model<any>): Promise<void> {
  const apiKey = await getApiKeyForModel(model);
  if (!apiKey) {
    throw new Error(`No API key for ${model.provider}/${model.id}`);
  }
  
  this.agent.setModel(model);
  this.sessionManager.saveModelChange(model.provider, model.id);
  this.settingsManager.setDefaultModelAndProvider(model.provider, model.id);
}

/**
 * Cycle to next model. Uses scoped models if available.
 * Returns null if only one model available.
 */
async cycleModel(): Promise<ModelCycleResult | null> {
  if (this.scopedModels.length > 0) {
    return this.cycleScopedModel();
  } else {
    return this.cycleAvailableModel();
  }
}

private async cycleScopedModel(): Promise<ModelCycleResult | null> {
  if (this.scopedModels.length <= 1) return null;
  
  const currentModel = this.model;
  let currentIndex = this.scopedModels.findIndex(
    (sm) => sm.model.id === currentModel?.id && sm.model.provider === currentModel?.provider
  );
  
  if (currentIndex === -1) currentIndex = 0;
  const nextIndex = (currentIndex + 1) % this.scopedModels.length;
  const next = this.scopedModels[nextIndex];
  
  // Validate API key
  const apiKey = await getApiKeyForModel(next.model);
  if (!apiKey) {
    throw new Error(`No API key for ${next.model.provider}/${next.model.id}`);
  }
  
  // Apply model
  this.agent.setModel(next.model);
  this.sessionManager.saveModelChange(next.model.provider, next.model.id);
  this.settingsManager.setDefaultModelAndProvider(next.model.provider, next.model.id);
  
  // Apply thinking level (silently use "off" if not supported)
  const effectiveThinking = next.model.reasoning ? next.thinkingLevel : "off";
  this.agent.setThinkingLevel(effectiveThinking);
  this.sessionManager.saveThinkingLevelChange(effectiveThinking);
  this.settingsManager.setDefaultThinkingLevel(effectiveThinking);
  
  return { model: next.model, thinkingLevel: effectiveThinking, isScoped: true };
}

private async cycleAvailableModel(): Promise<ModelCycleResult | null> {
  const { models: availableModels, error } = await getAvailableModels();
  if (error) throw new Error(`Failed to load models: ${error}`);
  if (availableModels.length <= 1) return null;
  
  const currentModel = this.model;
  let currentIndex = availableModels.findIndex(
    (m) => m.id === currentModel?.id && m.provider === currentModel?.provider
  );
  
  if (currentIndex === -1) currentIndex = 0;
  const nextIndex = (currentIndex + 1) % availableModels.length;
  const nextModel = availableModels[nextIndex];
  
  const apiKey = await getApiKeyForModel(nextModel);
  if (!apiKey) {
    throw new Error(`No API key for ${nextModel.provider}/${nextModel.id}`);
  }
  
  this.agent.setModel(nextModel);
  this.sessionManager.saveModelChange(nextModel.provider, nextModel.id);
  this.settingsManager.setDefaultModelAndProvider(nextModel.provider, nextModel.id);
  
  return { model: nextModel, thinkingLevel: this.thinkingLevel, isScoped: false };
}

/**
 * Get all available models with valid API keys.
 */
async getAvailableModels(): Promise<Model<any>[]> {
  const { models, error } = await getAvailableModels();
  if (error) throw new Error(error);
  return models;
}

Verification:

npm run check passes

Add ModelCycleResult interface
Add setModel() method
Add cycleModel() method with scoped/available variants
Add getAvailableModels() method
Verify with npm run check

WP6: AgentSession - Thinking Level Management

Add setThinkingLevel(), cycleThinkingLevel(), supportsThinking() methods.

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: cycleThinkingLevel() method (lines ~940-970)

Implementation:

// Add to AgentSession class

/**
 * Set thinking level. Silently uses "off" if model doesn't support it.
 * Saves to session and settings.
 */
setThinkingLevel(level: ThinkingLevel): void {
  const effectiveLevel = this.supportsThinking() ? level : "off";
  this.agent.setThinkingLevel(effectiveLevel);
  this.sessionManager.saveThinkingLevelChange(effectiveLevel);
  this.settingsManager.setDefaultThinkingLevel(effectiveLevel);
}

/**
 * Cycle to next thinking level.
 * Returns new level, or null if model doesn't support thinking.
 */
cycleThinkingLevel(): ThinkingLevel | null {
  if (!this.supportsThinking()) return null;
  
  const modelId = this.model?.id || "";
  const supportsXhigh = modelId.includes("codex-max");
  const levels: ThinkingLevel[] = supportsXhigh
    ? ["off", "minimal", "low", "medium", "high", "xhigh"]
    : ["off", "minimal", "low", "medium", "high"];
  
  const currentIndex = levels.indexOf(this.thinkingLevel);
  const nextIndex = (currentIndex + 1) % levels.length;
  const nextLevel = levels[nextIndex];
  
  this.setThinkingLevel(nextLevel);
  return nextLevel;
}

/**
 * Check if current model supports thinking.
 */
supportsThinking(): boolean {
  return !!this.model?.reasoning;
}

Verification:

npm run check passes

Add setThinkingLevel() method
Add cycleThinkingLevel() method
Add supportsThinking() method
Add setQueueMode() method and queueMode getter (see below)
Verify with npm run check

Queue mode (add to same WP):

// Add to AgentSession class

get queueMode(): QueueMode {
  return this.agent.getQueueMode();
}

/**
 * Set message queue mode. Saves to settings.
 */
setQueueMode(mode: QueueMode): void {
  this.agent.setQueueMode(mode);
  this.settingsManager.setQueueMode(mode);
}

WP7: AgentSession - Compaction

Add compact(), abortCompaction(), checkAutoCompaction(), autoCompactionEnabled methods.

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: executeCompaction() (lines ~2280-2370)
src/tui/tui-renderer.ts: checkAutoCompaction() (lines ~495-525)
src/main.ts: runRpcMode() auto-compaction logic (lines ~730-770)

Implementation:

// Add to AgentSession class

export interface CompactionResult {
  tokensBefore: number;
  tokensAfter: number;
  summary: string;
}

private compactionAbortController: AbortController | null = null;

/**
 * Manually compact the session context.
 * Aborts current agent operation first.
 */
async compact(customInstructions?: string): Promise<CompactionResult> {
  // Abort any running operation
  this.unsubscribeAll();
  await this.abort();
  
  // Create abort controller
  this.compactionAbortController = new AbortController();
  
  try {
    const apiKey = await getApiKeyForModel(this.model!);
    if (!apiKey) {
      throw new Error(`No API key for ${this.model!.provider}`);
    }
    
    const entries = this.sessionManager.loadEntries();
    const settings = this.settingsManager.getCompactionSettings();
    const compactionEntry = await compact(
      entries,
      this.model!,
      settings,
      apiKey,
      this.compactionAbortController.signal,
      customInstructions,
    );
    
    if (this.compactionAbortController.signal.aborted) {
      throw new Error("Compaction cancelled");
    }
    
    // Save and reload
    this.sessionManager.saveCompaction(compactionEntry);
    const loaded = loadSessionFromEntries(this.sessionManager.loadEntries());
    this.agent.replaceMessages(loaded.messages);
    
    return {
      tokensBefore: compactionEntry.tokensBefore,
      tokensAfter: compactionEntry.tokensAfter,
      summary: compactionEntry.summary,
    };
  } finally {
    this.compactionAbortController = null;
    // Note: caller needs to re-subscribe after compaction
  }
}

/**
 * Cancel in-progress compaction.
 */
abortCompaction(): void {
  this.compactionAbortController?.abort();
}

/**
 * Check if auto-compaction should run, and run if so.
 * Returns result if compaction occurred, null otherwise.
 */
async checkAutoCompaction(): Promise<CompactionResult | null> {
  const settings = this.settingsManager.getCompactionSettings();
  if (!settings.enabled) return null;
  
  // Get last non-aborted assistant message
  const messages = this.messages;
  let lastAssistant: AssistantMessage | null = null;
  for (let i = messages.length - 1; i >= 0; i--) {
    const msg = messages[i];
    if (msg.role === "assistant") {
      const assistantMsg = msg as AssistantMessage;
      if (assistantMsg.stopReason !== "aborted") {
        lastAssistant = assistantMsg;
        break;
      }
    }
  }
  if (!lastAssistant) return null;
  
  const contextTokens = calculateContextTokens(lastAssistant.usage);
  const contextWindow = this.model?.contextWindow ?? 0;
  
  if (!shouldCompact(contextTokens, contextWindow, settings)) return null;
  
  // Perform auto-compaction (don't abort current operation for auto)
  try {
    const apiKey = await getApiKeyForModel(this.model!);
    if (!apiKey) return null;
    
    const entries = this.sessionManager.loadEntries();
    const compactionEntry = await compact(entries, this.model!, settings, apiKey);
    
    this.sessionManager.saveCompaction(compactionEntry);
    const loaded = loadSessionFromEntries(this.sessionManager.loadEntries());
    this.agent.replaceMessages(loaded.messages);
    
    return {
      tokensBefore: compactionEntry.tokensBefore,
      tokensAfter: compactionEntry.tokensAfter,
      summary: compactionEntry.summary,
    };
  } catch {
    return null; // Silently fail auto-compaction
  }
}

/**
 * Toggle auto-compaction setting.
 */
setAutoCompactionEnabled(enabled: boolean): void {
  this.settingsManager.setCompactionEnabled(enabled);
}

get autoCompactionEnabled(): boolean {
  return this.settingsManager.getCompactionEnabled();
}

Verification:

npm run check passes

Add CompactionResult interface
Add compact() method
Add abortCompaction() method
Add checkAutoCompaction() method
Add setAutoCompactionEnabled() and getter
Verify with npm run check

WP8: AgentSession - Bash Execution

Add executeBash(), abortBash(), isBashRunning using the bash-executor module.

Files to modify:

src/core/agent-session.ts

Dependencies: WP1 (bash-executor.ts)

Implementation:

// Add to AgentSession class

import { executeBash as executeBashCommand, type BashResult } from "./bash-executor.js";
import type { BashExecutionMessage } from "../messages.js";

private bashAbortController: AbortController | null = null;

/**
 * Execute a bash command. Adds result to agent context and session.
 */
async executeBash(command: string, onChunk?: (chunk: string) => void): Promise<BashResult> {
  this.bashAbortController = new AbortController();
  
  try {
    const result = await executeBashCommand(command, {
      onChunk,
      signal: this.bashAbortController.signal,
    });
    
    // Create and save message
    const bashMessage: BashExecutionMessage = {
      role: "bashExecution",
      command,
      output: result.output,
      exitCode: result.exitCode,
      cancelled: result.cancelled,
      truncated: result.truncated,
      fullOutputPath: result.fullOutputPath,
      timestamp: Date.now(),
    };
    
    this.agent.appendMessage(bashMessage);
    this.sessionManager.saveMessage(bashMessage);
    
    // Initialize session if needed
    if (this.sessionManager.shouldInitializeSession(this.agent.state.messages)) {
      this.sessionManager.startSession(this.agent.state);
    }
    
    return result;
  } finally {
    this.bashAbortController = null;
  }
}

/**
 * Cancel running bash command.
 */
abortBash(): void {
  this.bashAbortController?.abort();
}

get isBashRunning(): boolean {
  return this.bashAbortController !== null;
}

Verification:

npm run check passes

Add bash execution methods using bash-executor module
Verify with npm run check

WP9: AgentSession - Session Management

Add switchSession(), branch(), getUserMessagesForBranching(), getSessionStats(), exportToHtml().

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: handleResumeSession() (lines ~1650-1710)
src/tui/tui-renderer.ts: showUserMessageSelector() branch logic (lines ~1560-1600)
src/tui/tui-renderer.ts: handleSessionCommand() (lines ~1870-1930)

Implementation:

// Add to AgentSession class

export interface SessionStats {
  sessionFile: string;
  sessionId: string;
  userMessages: number;
  assistantMessages: number;
  toolCalls: number;
  toolResults: number;
  totalMessages: number;
  tokens: {
    input: number;
    output: number;
    cacheRead: number;
    cacheWrite: number;
    total: number;
  };
  cost: number;
}

/**
 * Switch to a different session file.
 * Aborts current operation, loads messages, restores model/thinking.
 */
async switchSession(sessionPath: string): Promise<void> {
  this.unsubscribeAll();
  await this.abort();
  this.queuedMessages = [];
  
  this.sessionManager.setSessionFile(sessionPath);
  const loaded = loadSessionFromEntries(this.sessionManager.loadEntries());
  this.agent.replaceMessages(loaded.messages);
  
  // Restore model
  const savedModel = this.sessionManager.loadModel();
  if (savedModel) {
    const availableModels = (await getAvailableModels()).models;
    const match = availableModels.find(
      (m) => m.provider === savedModel.provider && m.id === savedModel.modelId
    );
    if (match) {
      this.agent.setModel(match);
    }
  }
  
  // Restore thinking level
  const savedThinking = this.sessionManager.loadThinkingLevel();
  if (savedThinking) {
    this.agent.setThinkingLevel(savedThinking as ThinkingLevel);
  }
  
  // Note: caller needs to re-subscribe after switch
}

/**
 * Create a branch from a specific entry index.
 * Returns the text of the selected user message (for editor pre-fill).
 */
branch(entryIndex: number): string {
  const entries = this.sessionManager.loadEntries();
  const selectedEntry = entries[entryIndex];
  
  if (selectedEntry.type !== "message" || selectedEntry.message.role !== "user") {
    throw new Error("Invalid entry index for branching");
  }
  
  const selectedText = this.extractUserMessageText(selectedEntry.message.content);
  
  // Create branched session
  const newSessionFile = this.sessionManager.createBranchedSessionFromEntries(entries, entryIndex);
  this.sessionManager.setSessionFile(newSessionFile);
  
  // Reload
  const loaded = loadSessionFromEntries(this.sessionManager.loadEntries());
  this.agent.replaceMessages(loaded.messages);
  
  return selectedText;
}

/**
 * Get all user messages from session for branch selector.
 */
getUserMessagesForBranching(): Array<{ entryIndex: number; text: string }> {
  const entries = this.sessionManager.loadEntries();
  const result: Array<{ entryIndex: number; text: string }> = [];
  
  for (let i = 0; i < entries.length; i++) {
    const entry = entries[i];
    if (entry.type !== "message") continue;
    if (entry.message.role !== "user") continue;
    
    const text = this.extractUserMessageText(entry.message.content);
    if (text) {
      result.push({ entryIndex: i, text });
    }
  }
  
  return result;
}

private extractUserMessageText(content: string | Array<{ type: string; text?: string }>): string {
  if (typeof content === "string") return content;
  if (Array.isArray(content)) {
    return content
      .filter((c): c is { type: "text"; text: string } => c.type === "text")
      .map((c) => c.text)
      .join("");
  }
  return "";
}

/**
 * Get session statistics.
 */
getSessionStats(): SessionStats {
  const state = this.state;
  const userMessages = state.messages.filter((m) => m.role === "user").length;
  const assistantMessages = state.messages.filter((m) => m.role === "assistant").length;
  const toolResults = state.messages.filter((m) => m.role === "toolResult").length;
  
  let toolCalls = 0;
  let totalInput = 0;
  let totalOutput = 0;
  let totalCacheRead = 0;
  let totalCacheWrite = 0;
  let totalCost = 0;
  
  for (const message of state.messages) {
    if (message.role === "assistant") {
      const assistantMsg = message as AssistantMessage;
      toolCalls += assistantMsg.content.filter((c) => c.type === "toolCall").length;
      totalInput += assistantMsg.usage.input;
      totalOutput += assistantMsg.usage.output;
      totalCacheRead += assistantMsg.usage.cacheRead;
      totalCacheWrite += assistantMsg.usage.cacheWrite;
      totalCost += assistantMsg.usage.cost.total;
    }
  }
  
  return {
    sessionFile: this.sessionFile,
    sessionId: this.sessionId,
    userMessages,
    assistantMessages,
    toolCalls,
    toolResults,
    totalMessages: state.messages.length,
    tokens: {
      input: totalInput,
      output: totalOutput,
      cacheRead: totalCacheRead,
      cacheWrite: totalCacheWrite,
      total: totalInput + totalOutput + totalCacheRead + totalCacheWrite,
    },
    cost: totalCost,
  };
}

/**
 * Export session to HTML.
 */
exportToHtml(outputPath?: string): string {
  return exportSessionToHtml(this.sessionManager, this.state, outputPath);
}

Verification:

npm run check passes

Add SessionStats interface
Add switchSession() method
Add branch() method
Add getUserMessagesForBranching() method
Add getSessionStats() method
Add exportToHtml() method
Verify with npm run check

WP10: AgentSession - Utility Methods

Add getLastAssistantText() and any remaining utilities.

Files to modify:

src/core/agent-session.ts

Extract from:

src/tui/tui-renderer.ts: handleCopyCommand() (lines ~1840-1870)

Implementation:

// Add to AgentSession class

/**
 * Get text content of last assistant message (for /copy).
 * Returns null if no assistant message exists.
 */
getLastAssistantText(): string | null {
  const lastAssistant = this.messages
    .slice()
    .reverse()
    .find((m) => m.role === "assistant");
  
  if (!lastAssistant) return null;
  
  let text = "";
  for (const content of lastAssistant.content) {
    if (content.type === "text") {
      text += content.text;
    }
  }
  
  return text.trim() || null;
}

/**
 * Get queued message count (for UI display).
 */
get queuedMessageCount(): number {
  return this.queuedMessages.length;
}

/**
 * Get queued messages (for display, not modification).
 */
getQueuedMessages(): readonly string[] {
  return this.queuedMessages;
}

Verification:

npm run check passes

Add getLastAssistantText() method
Add queuedMessageCount getter (done in WP4)
Add getQueuedMessages() method (done in WP4)
Verify with npm run check

WP11: Create print-mode.ts

Extract single-shot mode into its own module using AgentSession.

Files to create:

src/modes/print-mode.ts

Extract from:

src/main.ts: runSingleShotMode() function (lines ~615-640)

Implementation:

// src/modes/print-mode.ts

import type { Attachment } from "@mariozechner/pi-agent-core";
import type { AssistantMessage } from "@mariozechner/pi-ai";
import type { AgentSession } from "../core/agent-session.js";

export async function runPrintMode(
  session: AgentSession,
  mode: "text" | "json",
  messages: string[],
  initialMessage?: string,
  initialAttachments?: Attachment[],
): Promise<void> {
  
  if (mode === "json") {
    // Output all events as JSON
    session.subscribe((event) => {
      console.log(JSON.stringify(event));
    });
  }

  // Send initial message with attachments
  if (initialMessage) {
    await session.prompt(initialMessage, { attachments: initialAttachments });
  }

  // Send remaining messages
  for (const message of messages) {
    await session.prompt(message);
  }

  // In text mode, output final response
  if (mode === "text") {
    const state = session.state;
    const lastMessage = state.messages[state.messages.length - 1];
    
    if (lastMessage?.role === "assistant") {
      const assistantMsg = lastMessage as AssistantMessage;
      
      // Check for error/aborted
      if (assistantMsg.stopReason === "error" || assistantMsg.stopReason === "aborted") {
        console.error(assistantMsg.errorMessage || `Request ${assistantMsg.stopReason}`);
        process.exit(1);
      }
      
      // Output text content
      for (const content of assistantMsg.content) {
        if (content.type === "text") {
          console.log(content.text);
        }
      }
    }
  }
}

Verification:

npm run check passes
Manual test: pi -p "echo hello" still works

Create src/modes/print-mode.ts
Verify with npm run check

WP12: Create rpc-mode.ts

Extract RPC mode into its own module using AgentSession.

Files to create:

src/modes/rpc-mode.ts

Extract from:

src/main.ts: runRpcMode() function (lines ~700-800)

Implementation:

// src/modes/rpc-mode.ts

import * as readline from "readline";
import type { AgentSession } from "../core/agent-session.js";

export async function runRpcMode(session: AgentSession): Promise<never> {
  // Output all events as JSON
  session.subscribe((event) => {
    console.log(JSON.stringify(event));
    
    // Emit auto-compaction events
    // (checkAutoCompaction is called internally by AgentSession after assistant messages)
  });

  // Listen for JSON input
  const rl = readline.createInterface({
    input: process.stdin,
    output: process.stdout,
    terminal: false,
  });

  rl.on("line", async (line: string) => {
    try {
      const input = JSON.parse(line);

      switch (input.type) {
        case "prompt":
          if (input.message) {
            await session.prompt(input.message, { 
              attachments: input.attachments,
              expandSlashCommands: false, // RPC mode doesn't expand slash commands
            });
          }
          break;

        case "abort":
          await session.abort();
          break;

        case "compact":
          try {
            const result = await session.compact(input.customInstructions);
            console.log(JSON.stringify({ type: "compaction", ...result }));
          } catch (error: any) {
            console.log(JSON.stringify({ type: "error", error: `Compaction failed: ${error.message}` }));
          }
          break;

        case "bash":
          if (input.command) {
            try {
              const result = await session.executeBash(input.command);
              console.log(JSON.stringify({ type: "bash_end", message: result }));
            } catch (error: any) {
              console.log(JSON.stringify({ type: "error", error: `Bash failed: ${error.message}` }));
            }
          }
          break;

        default:
          console.log(JSON.stringify({ type: "error", error: `Unknown command: ${input.type}` }));
      }
    } catch (error: any) {
      console.log(JSON.stringify({ type: "error", error: error.message }));
    }
  });

  // Keep process alive forever
  return new Promise(() => {});
}

Verification:

npm run check passes
Manual test: RPC mode still works (if you have a way to test it)

Create src/modes/rpc-mode.ts
Verify with npm run check

WP13: Create modes/index.ts barrel export

Create barrel export for all modes.

Files to create:

src/modes/index.ts

Implementation:

// src/modes/index.ts
export { runPrintMode } from "./print-mode.js";
export { runRpcMode } from "./rpc-mode.js";
// InteractiveMode will be added later

Create src/modes/index.ts
Verify with npm run check

WP14: Create main-new.ts using AgentSession and new modes

Create a new main file that uses AgentSession and the new mode modules. Old main.ts is kept for reference/comparison.

Files to create:

src/main-new.ts (copy from main.ts, then modify)
src/cli-new.ts (copy from cli.ts, point to main-new.ts)

Changes to main-new.ts:

Remove runSingleShotMode() function (use print-mode.ts)
Remove runRpcMode() function (use rpc-mode.ts)
Remove executeRpcBashCommand() function (use bash-executor.ts)
Create AgentSession instance after agent setup
Pass AgentSession to mode functions

Key changes in main():

// After agent creation, create AgentSession
const session = new AgentSession({
  agent,
  sessionManager,
  settingsManager,
  scopedModels,
  fileCommands: loadSlashCommands(),
});

// Route to modes
if (mode === "rpc") {
  await runRpcMode(session);
} else if (isInteractive) {
  // For now, still use TuiRenderer directly (will refactor in WP15+)
  await runInteractiveMode(agent, sessionManager, ...);
} else {
  await runPrintMode(session, mode, parsed.messages, initialMessage, initialAttachments);
}

cli-new.ts:

#!/usr/bin/env node
import { main } from "./main-new.js";
main(process.argv.slice(2));

Testing the new implementation:

# Run new implementation directly
npx tsx src/cli-new.ts -p "hello"
npx tsx src/cli-new.ts --mode json "hello"
npx tsx src/cli-new.ts  # interactive mode

Verification:

npm run check passes
Manual test: npx tsx src/cli-new.ts -p "hello" works
Manual test: npx tsx src/cli-new.ts --mode json "hello" works
Manual test: npx tsx src/cli-new.ts --mode rpc works

Copy main.ts to main-new.ts
Remove runSingleShotMode() from main-new.ts
Remove runRpcMode() from main-new.ts
Remove executeRpcBashCommand() from main-new.ts
Import and use runPrintMode from modes
Import and use runRpcMode from modes
Create AgentSession in main()
Update mode routing to use new functions
Create cli-new.ts
Verify with npm run check
Manual test all three modes via cli-new.ts

WP15: Create InteractiveMode using AgentSession

Create a new interactive mode class that uses AgentSession. Old tui-renderer.ts is kept for reference.

Files to create:

src/modes/interactive/interactive-mode.ts (based on tui-renderer.ts)

This is the largest change. Strategy:

Copy tui-renderer.ts to new location
Rename class from TuiRenderer to InteractiveMode
Change constructor to accept AgentSession instead of separate agent/sessionManager/settingsManager
Replace all this.agent.* calls with this.session.agent.* or appropriate AgentSession methods
Replace all this.sessionManager.* calls with AgentSession methods
Replace all this.settingsManager.* calls with AgentSession methods where applicable
Remove duplicated logic that now lives in AgentSession

Key replacements:

Old	New
`this.agent.prompt()`	`this.session.prompt()`
`this.agent.abort()`	`this.session.abort()`
`this.sessionManager.saveMessage()`	(handled internally by AgentSession.subscribe)
`this.cycleThinkingLevel()`	`this.session.cycleThinkingLevel()`
`this.cycleModel()`	`this.session.cycleModel()`
`this.executeBashCommand()`	`this.session.executeBash()`
`this.executeCompaction()`	`this.session.compact()`
`this.checkAutoCompaction()`	(handled internally by AgentSession)
`this.handleClearCommand()` reset logic	`this.session.reset()`
`this.handleResumeSession()`	`this.session.switchSession()`

Constructor change:

// Old
constructor(
  agent: Agent,
  sessionManager: SessionManager,
  settingsManager: SettingsManager,
  version: string,
  ...
)

// New  
constructor(
  session: AgentSession,
  version: string,
  ...
)

Verification:

npm run check passes
Manual test via cli-new.ts: Full interactive mode works
Manual test: All slash commands work
Manual test: All hotkeys work
Manual test: Bash execution works
Manual test: Model/thinking cycling works

Create src/modes/interactive/ directory
Copy tui-renderer.ts to interactive-mode.ts
Rename class to InteractiveMode
Change constructor to accept AgentSession
Update all agent access to go through session
Remove subscribeToAgent() method (use session.subscribe)
Remove checkAutoCompaction() method (handled by session)
Update cycleThinkingLevel() to use session method
Update cycleModel() to use session method
Update bash execution to use session.executeBash()
Update compaction to use session.compact()
Update reset logic to use session.reset()
Update session switching to use session.switchSession()
Update branch logic to use session.branch()
Remove all direct sessionManager access
Update imports to point to ../../tui/ for components (keep old components in place for now)
Update modes/index.ts to export InteractiveMode
Verify with npm run check
Manual test interactive mode via cli-new.ts

WP16: Update main-new.ts runInteractiveMode to use InteractiveMode

Update runInteractiveMode in main-new.ts to use the new InteractiveMode class.

Files to modify:

src/main-new.ts

Changes:

import { InteractiveMode } from "./modes/interactive/interactive-mode.js";

async function runInteractiveMode(
  session: AgentSession,
  version: string,
  changelogMarkdown: string | null,
  collapseChangelog: boolean,
  modelFallbackMessage: string | null,
  versionCheckPromise: Promise<string | null>,
  initialMessages: string[],
  initialMessage?: string,
  initialAttachments?: Attachment[],
  fdPath: string | null,
): Promise<void> {
  const mode = new InteractiveMode(
    session,
    version,
    changelogMarkdown,
    collapseChangelog,
    fdPath,
  );
  // ... rest stays similar
}

Verification:

npm run check passes
Manual test via cli-new.ts: Interactive mode works

Update runInteractiveMode() in main-new.ts
Update InteractiveMode instantiation
Verify with npm run check

WP17: (OPTIONAL) Move TUI components to modes/interactive/

Move TUI-specific components to the interactive mode directory. This is optional cleanup - can be skipped if too disruptive.

Note: The old src/tui/ directory is kept. We just create copies/moves as needed. For now, InteractiveMode can import from ../../tui/ to reuse existing components.

Files to potentially move (if doing this WP):

src/tui/assistant-message.ts → src/modes/interactive/components/
src/tui/bash-execution.ts → src/modes/interactive/components/
etc.

Skip this WP for now - focus on getting the new architecture working first. The component organization can be cleaned up later.

SKIPPED (optional cleanup for later)

WP19: Extract setup logic from main.ts

Create setup.ts with model resolution, system prompt building, etc.

Files to create:

src/core/setup.ts

Extract from main.ts:

buildSystemPrompt() function
loadProjectContextFiles() function
loadContextFileFromDir() function
resolveModelScope() function
Model resolution logic (the priority system)
Session loading/restoration logic

Implementation:

// src/core/setup.ts

export interface SetupOptions {
  provider?: string;
  model?: string;
  apiKey?: string;
  systemPrompt?: string;
  appendSystemPrompt?: string;
  thinking?: ThinkingLevel;
  continue?: boolean;
  resume?: boolean;
  models?: string[];
  tools?: ToolName[];
  sessionManager: SessionManager;
  settingsManager: SettingsManager;
}

export interface SetupResult {
  agent: Agent;
  initialModel: Model<any> | null;
  initialThinking: ThinkingLevel;
  scopedModels: Array<{ model: Model<any>; thinkingLevel: ThinkingLevel }>;
  modelFallbackMessage: string | null;
}

export async function setupAgent(options: SetupOptions): Promise<SetupResult>;

export function buildSystemPrompt(
  customPrompt?: string, 
  selectedTools?: ToolName[], 
  appendSystemPrompt?: string
): string;

export function loadProjectContextFiles(): Array<{ path: string; content: string }>;

export async function resolveModelScope(
  patterns: string[]
): Promise<Array<{ model: Model<any>; thinkingLevel: ThinkingLevel }>>;

Verification:

npm run check passes
All modes still work

Create src/core/setup.ts
Move buildSystemPrompt() from main.ts
Move loadProjectContextFiles() from main.ts
Move loadContextFileFromDir() from main.ts
Move resolveModelScope() from main.ts
Create setupAgent() function
Update main.ts to use setup.ts
Verify with npm run check

WP20: Final cleanup and documentation

Clean up main.ts, add documentation, verify everything works.

Tasks:

Remove any dead code from main.ts
Ensure main.ts is ~200-300 lines (just arg parsing + routing)
Add JSDoc comments to AgentSession public methods
Update README if needed
Final manual testing of all features

Verification:

npm run check passes
All three modes work
All slash commands work
All hotkeys work
Session persistence works
Compaction works
Bash execution works
Model/thinking cycling works

Remove dead code from main.ts
Add JSDoc to AgentSession
Final testing
Update README if needed

Testing Checklist (E2E)

After refactoring is complete, verify these scenarios:

Interactive Mode

Start fresh session: pi
Continue session: pi -c
Resume session: pi -r
Initial message: pi "hello"
File attachment: pi @file.txt "summarize"
Model cycling: Ctrl+P
Thinking cycling: Shift+Tab
Tool expansion: Ctrl+O
Thinking toggle: Ctrl+T
Abort: Esc during streaming
Clear: Ctrl+C twice to exit
Bash command: !ls -la
Bash cancel: Esc during bash
/thinking command
/model command
/export command
/copy command
/session command
/changelog command
/branch command
/login and /logout commands
/queue command
/theme command
/clear command
/compact command
/autocompact command
/resume command
Message queuing while streaming

Print Mode

Basic: pi -p "hello"
JSON: pi --mode json "hello"
Multiple messages: pi -p "first" "second"
File attachment: pi -p @file.txt "summarize"

RPC Mode

Start: pi --mode rpc
Send prompt via JSON
Abort via JSON
Compact via JSON
Bash via JSON

Notes

This refactoring should be done incrementally, testing after each work package
If a WP introduces regressions, fix them before moving to the next
The most risky WP is WP15 (updating TuiRenderer) - take extra care there
Consider creating git commits after each major WP for easy rollback

49 KiB Raw Blame History

Coding Agent Refactoring Plan

Status

Strategy: Keep Old Code for Reference

Goals

Architecture Overview

Current State (Problems)

Target State

AgentSession API

Work Packages

WP1: Create bash-executor.ts

WP2: Create agent-session.ts (Core Structure)

WP3: AgentSession - Event Subscription + Session Persistence

WP4: AgentSession - Prompting Methods

WP5: AgentSession - Model Management

WP6: AgentSession - Thinking Level Management

WP7: AgentSession - Compaction

WP8: AgentSession - Bash Execution

WP9: AgentSession - Session Management

WP10: AgentSession - Utility Methods

WP11: Create print-mode.ts

WP12: Create rpc-mode.ts

WP13: Create modes/index.ts barrel export

WP14: Create main-new.ts using AgentSession and new modes

WP15: Create InteractiveMode using AgentSession

WP16: Update main-new.ts runInteractiveMode to use InteractiveMode

WP17: (OPTIONAL) Move TUI components to modes/interactive/

WP19: Extract setup logic from main.ts

WP20: Final cleanup and documentation

Testing Checklist (E2E)

Interactive Mode

Print Mode

RPC Mode

Notes

49 KiB

Raw Blame History