The codex module enhances OpenAI's codex terminal agent, enabling semantic code search, automated patch generation, and connection to various AI model providers within a secure, sandboxed environment.
When to Activate
- When the user wants to start a Codex session with a specific provider (e.g., DeepSeek, Kimi, Zhipu, Doubao, SiliconFlow).
- When the user needs to execute shell commands or generate code based on natural language descriptions.
- When the user requires a sandboxed execution environment (
read-only,workspace-write) for safety. - When the user wants to apply generated diffs to their Git worktree.
- When the user needs to inject specific "skills" into the Codex agent's environment.
Core Principles & Rules
- Sandboxing: Always encourage using appropriate sandbox strategies (
--sandbox) to prevent unintended system changes. - Automation Levels: Use
--full-autofor a balance between speed and safety (approves on failure). - Provider Switching: Use subcommands like
ds(DeepSeek),kimi,zhipu,doubao, orsili(SiliconFlow) to launch Codex with these providers. - Non-Interactive Execution: Use
execorefor quick, one-off commands or code generation tasks.
Additional Scenarios
- Web Search: Enable real-time web search capabilities using the
--searchflag for up-to-date information. - Git Integration: Quickly apply the latest generated diff from the agent using
x codex apply. - Local OSS Models: Connect to local Ollama services by using the
--ossflag.
Patterns & Examples
Run Natural Language Command
# Execute a command described in English
x codex e "List all files larger than 10MB in the current directory"
Start with Third-party Providers
# Launch Codex using DeepSeek with read-only sandbox protection
x codex ds --sandbox read-only
# Launch Codex using SiliconFlow
x codex sili
Apply Latest Patch
# Apply the latest generated code patch to the current Git repository
x codex apply
Inject Skills
# Manage and inject custom skills into the Codex agent environment
x codex skill
Checklist
- Confirm the desired sandbox level (
read-only,workspace-write, etc.). - Verify if a specific model provider (DeepSeek, Kimi, etc.) is needed.
- Ensure the user is aware of the risks when using
dangerously-bypass-approvals-and-sandbox.