feat: implement modular provider architecture with Codex CLI support

Implements a flexible provider pattern that supports both Claude Agent SDK
and OpenAI Codex CLI, enabling future expansion to other AI providers
(Cursor, OpenCode, etc.) with minimal changes.

## Architecture Changes

### New Provider System
- Created provider abstraction layer with BaseProvider interface
- Model-based routing: model prefix determines provider
  - `gpt-*`, `o*` → CodexProvider (subprocess CLI)
  - `claude-*`, `opus/sonnet/haiku` → ClaudeProvider (SDK)
- Providers implement common ExecuteOptions interface

### New Files Created
- `providers/types.ts` - Shared interfaces (ExecuteOptions, ProviderMessage, etc.)
- `providers/base-provider.ts` - Abstract base class
- `providers/claude-provider.ts` - Claude Agent SDK wrapper
- `providers/codex-provider.ts` - Codex CLI subprocess executor
- `providers/codex-cli-detector.ts` - Installation & auth detection
- `providers/codex-config-manager.ts` - TOML config management
- `providers/provider-factory.ts` - Model-based provider routing
- `lib/subprocess-manager.ts` - Reusable subprocess utilities

## Features Implemented

### Codex CLI Integration
- Spawns Codex CLI as subprocess with JSONL output
- Converts Codex events to Claude SDK-compatible format
- Supports both `codex login` and OPENAI_API_KEY auth methods
- Handles: reasoning, messages, commands, todos, file changes
- Extracts text from content blocks for non-vision CLI

### Conversation History
- Added conversationHistory support to ExecuteOptions
- ClaudeProvider: yields previous messages to SDK
- CodexProvider: prepends history as text context
- Follow-up prompts maintain full conversation context

### Image Upload Support
- Images embedded as base64 for vision models
- Image paths appended to prompt text for Read tool access
- Auto-mode: copies images to feature folder
- Follow-up: combines original + new images
- Updates feature.json with image metadata

### Session Model Persistence
- Added `model` field to Session and SessionMetadata
- Sessions remember model preference across interactions
- API endpoints accept model parameter
- Auto-mode respects feature's model setting

## Modified Files

### Services
- `agent-service.ts`:
  - Added conversation history building
  - Uses ProviderFactory instead of direct SDK calls
  - Appends image paths to prompts
  - Added model parameter and persistence

- `auto-mode-service.ts`:
  - Removed OpenAI model block restriction
  - Uses ProviderFactory for all models
  - Added image support in buildFeaturePrompt
  - Follow-up: loads context, copies images, updates feature.json
  - Returns to waiting_approval after follow-up

### Routes
- `agent.ts`: Added model parameter to /send endpoint
- `sessions.ts`: Added model field to create/update
- `models.ts`: Added Codex models (gpt-5.2, gpt-5.1-codex*)

### Configuration
- `.env.example`: Added OPENAI_API_KEY and CODEX_CLI_PATH
- `.gitignore`: Added provider-specific ignores

## Bug Fixes
- Fixed image path resolution (relative → absolute)
- Fixed Codex empty prompt when images attached
- Fixed follow-up status management (in_progress → waiting_approval)
- Fixed follow-up images not appearing in prompt text
- Removed OpenAI model restrictions in auto-mode

## Testing Notes
- Codex CLI authentication verified with both methods
- Image uploads work for both Claude (vision) and Codex (Read tool)
- Follow-up prompts maintain full context
- Conversation history persists across turns
- Model switching works per-session

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

apps/server/src/providers/base-provider.ts

@@ -0,0 +1,96 @@
+/**
+ * Abstract base class for AI model providers
+ */
+
+import type {
+  ProviderConfig,
+  ExecuteOptions,
+  ProviderMessage,
+  InstallationStatus,
+  ValidationResult,
+  ModelDefinition,
+} from "./types.js";
+
+/**
+ * Base provider class that all provider implementations must extend
+ */
+export abstract class BaseProvider {
+  protected config: ProviderConfig;
+  protected name: string;
+
+  constructor(config: ProviderConfig = {}) {
+    this.config = config;
+    this.name = this.getName();
+  }
+
+  /**
+   * Get the provider name (e.g., "claude", "codex", "cursor")
+   */
+  abstract getName(): string;
+
+  /**
+   * Execute a query and stream responses
+   * @param options Execution options
+   * @returns AsyncGenerator yielding provider messages
+   */
+  abstract executeQuery(
+    options: ExecuteOptions
+  ): AsyncGenerator<ProviderMessage>;
+
+  /**
+   * Detect if the provider is installed and configured
+   * @returns Installation status
+   */
+  abstract detectInstallation(): Promise<InstallationStatus>;
+
+  /**
+   * Get available models for this provider
+   * @returns Array of model definitions
+   */
+  abstract getAvailableModels(): ModelDefinition[];
+
+  /**
+   * Validate the provider configuration
+   * @returns Validation result
+   */
+  validateConfig(): ValidationResult {
+    const errors: string[] = [];
+    const warnings: string[] = [];
+
+    // Base validation (can be overridden)
+    if (!this.config) {
+      errors.push("Provider config is missing");
+    }
+
+    return {
+      valid: errors.length === 0,
+      errors,
+      warnings,
+    };
+  }
+
+  /**
+   * Check if the provider supports a specific feature
+   * @param feature Feature name (e.g., "vision", "tools", "mcp")
+   * @returns Whether the feature is supported
+   */
+  supportsFeature(feature: string): boolean {
+    // Default implementation - override in subclasses
+    const commonFeatures = ["tools", "text"];
+    return commonFeatures.includes(feature);
+  }
+
+  /**
+   * Get provider configuration
+   */
+  getConfig(): ProviderConfig {
+    return this.config;
+  }
+
+  /**
+   * Update provider configuration
+   */
+  setConfig(config: Partial<ProviderConfig>): void {
+    this.config = { ...this.config, ...config };
+  }
+}