Repository navigation
Rebuilt on the design from Anthropic's How we built Claude Code auto mode.
- Every permission ask goes to the classifier, and so do allowed shell commands that can run arbitrary code
- The classifier sees only user messages and tool calls, never agent prose or tool output
- Scripts the agent wrote and then runs are judged by their contents
- Subagents are judged against the real user's messages
- Prompt injection probe on tool output
- Hard and clearable block rules; extend with trustedEnvironment, blockRules, allowRules
- Default reviewer: DeepSeek V4.1 Flash at high reasoning; optional fast stage 1 via fastModel
- Any provider model: provider/model#variant
- 3 blocks in a row tells the agent to ask you; 20 pauses auto mode until your next message
Install: opencode plugin add github:anargia-pixels/opencode-automode-plugin#release