feat: add codex claude sdk

README.md

@@ -80,7 +80,7 @@ We also generate multimodal skin based on the rules of the same maze. Generated
 - [ ] Add Multimodal Environment Generation Pipelines.
 - [ ] Add Three-stage Verification Pipeline for both text and multimodal environments.
 - [ ] Add Learning Experiments Scripts.
-- [ ] Add Coding Agents Option: Codex, Gemini Cli.
+- [x] Add Coding Agents Option: Codex, Claude Code SDK.
 - [ ] Add 3D Environment Generation Pipelines.
 
 ## Repository Layout
@@ -129,6 +129,200 @@ python run_environment_skin_generation.py
 
 Cost summaries are automatically saved to `workspace/costs/`.
 
+## Coding Agents
+
+AutoEnv supports multiple coding agent backends for environment code generation and fixing. The code agent is used in the pipeline's CodeFixNode, LevelGenNode, and MaxRewardNode stages.
+
+### Backend Options
+
+| Backend   | Description                                                                  | Best For                            |
+| --------- | ---------------------------------------------------------------------------- | ----------------------------------- |
+| `miniswe` | Default [mini-swe-agent](https://github.com/SWE-agent/mini-swe-agent)        | General use, works with any LLM     |
+| `codex`   | OpenAI [Codex CLI](https://github.com/openai/codex)                          | OpenAI API users, fast execution    |
+| `claude`  | Anthropic [Claude Agent SDK](https://github.com/anthropics/claude-agent-sdk) | Anthropic API users, custom proxies |
+
+### Configuration
+
+Set the `code_agent_backend` field in `config/env_gen.yaml`:
+
+```yaml
+# Code agent backend: "miniswe" (default), "codex", "claude"
+code_agent_backend: "codex" # or "claude" or "miniswe"
+```
+
+### MiniSWE Agent (Default)
+
+The default agent using mini-swe-agent. Works with any LLM configured in `config/model_config.yaml`.
+
+```yaml
+# config/env_gen.yaml
+code_agent_backend: "miniswe"
+model: "gpt-4o" # or any configured LLM
+```
+
+No additional setup required beyond the standard model configuration.
+
+### Codex Agent
+
+Uses OpenAI's official Codex CLI for code generation.
+
+#### Prerequisites
+
+```bash
+# Install Codex CLI
+npm install -g @openai/codex
+# or
+brew install --cask codex
+
+# Authenticate (recommended)
+codex login
+
+# Verify installation
+codex whoami
+```
+
+#### Authentication Options
+
+**Option 1: CLI Login (Recommended)**
+
+```bash
+codex login  # Opens browser for OAuth
+```
+
+**Option 2: Environment Variable**
+
+```bash
+export OPENAI_API_KEY=your-api-key
+```
+
+**Option 3: Custom Base URL** (for proxies)
+
+```bash
+export OPENAI_API_KEY=your-api-key
+export OPENAI_BASE_URL=https://your-proxy.example.com/v1
+```
+
+#### Configuration
+
+```yaml
+# config/env_gen.yaml
+code_agent_backend: "codex"
+```
+
+### Claude Agent
+
+Uses Anthropic's Claude Agent SDK for code generation.
+
+#### Prerequisites
+
+```bash
+# Install Claude Agent SDK
+pip install claude-agent-sdk
+```
+
+The Python SDK authenticates via environment variables (see Authentication Options below).
+
+#### Authentication Options
+
+**Option 1: Environment Variables (Recommended)**
+
+```bash
+export ANTHROPIC_API_KEY=your-api-key
+```
+
+**Option 2: Custom Base URL** (for proxies)
+
+```bash
+export ANTHROPIC_API_KEY=your-api-key
+export ANTHROPIC_BASE_URL=https://your-proxy.example.com/api
+```
+
+#### Configuration
+
+```yaml
+# config/env_gen.yaml
+code_agent_backend: "claude"
+```
+
+### Generate Environment with Code Agent
+
+#### Quick Start
+
+```bash
+# 1. Copy example config
+cp config/env_gen_example.yaml config/env_gen.yaml
+
+# 2. Edit config (set code_agent_backend, theme, etc.)
+vim config/env_gen.yaml
+
+# 3. Run environment generation
+python run_environment_generation.py
+```
+
+#### Example Configuration
+
+```yaml
+# config/env_gen.yaml
+mode: "textual"
+model: "gpt-4o"
+concurrency: 1
+theme: "A strategic puzzle game with resource management"
+envs_root_path: "workspace/envs"
+code_agent_backend: "codex" # Use Codex for code generation
+```
+
+#### Background Execution (Recommended for Long Tasks)
+
+Code agents can take 10-30 minutes for complex environments. Run in background:
+
+```bash
+# Run in background with logging
+nohup python run_environment_generation.py > /tmp/autoenv_gen.log 2>&1 &
+
+# Monitor progress
+tail -f /tmp/autoenv_gen.log
+
+# Check if complete
+ls workspace/envs/*/done.txt
+```
+
+### Troubleshooting
+
+#### Codex CLI Issues
+
+```bash
+# Check if Codex is installed
+codex --version
+
+# Re-authenticate
+codex logout && codex login
+
+# Check current user
+codex whoami
+```
+
+#### Claude Agent Issues
+
+```bash
+# Check if Python SDK is installed
+pip show claude-agent-sdk
+
+# Verify API key
+echo $ANTHROPIC_API_KEY
+
+# Test import
+python -c "from claude_agent_sdk import query; print('SDK available')"
+```
+
+#### Timeout Issues
+
+For complex environments, increase timeout in `autoenv/coder.py`:
+
+```python
+# Current default: 900 seconds (15 minutes)
+agent = CodexAgent(timeout=1200)  # 20 minutes
+```
+
 ## Benchmarking AutoEnv-36
 
 Evaluate agents on the 36 benchmark environments (scores for all; cost only for LLM branch). See `benchmarks/README.md` for details.
@@ -160,9 +354,10 @@ Programmatic APIs are available in `benchmarks/api.py` (`benchmark_llms`, `bench
 
 ## Acknowledgements
 
-Thanks to 
-[mini-swe-agent](https://github.com/SWE-agent/mini-swe-agent), 
-[codex](https://github.com/openai/codex), 
+Thanks to
+[mini-swe-agent](https://github.com/SWE-agent/mini-swe-agent),
+[codex](https://github.com/openai/codex),
+[claude-agent-sdk](https://github.com/anthropics/claude-agent-sdk-python),
 [rembg](https://github.com/danielgatis/rembg),
 for providing basic support for this project!