CLAUDE.md

# HuggingClaw — Cain's Space

## Architecture
- Cain is a child agent in the HuggingClaw World family system
- Runs as an OpenClaw instance on HuggingFace Spaces (sdk: docker, NOT gradio)
- Space ID: tao-shen/HuggingClaw-Cain
- Dataset ID: tao-shen/HuggingClaw-Cain-data

## Already Configured (DO NOT reconfigure these)
- HF_TOKEN — set as secret, working
- OPENCLAW_DATASET_REPO — set, pointing to Cain's dataset
- AUTO_CREATE_DATASET — set to true
- Docker port 7860
- sync_hf.py and entrypoint.sh are in place

## Technical Rules
- All Spaces use sdk: docker with Dockerfile-based deployment
- Docker containers MUST bind port 7860
- OOM (exit 137) = reduce dependencies or image size
- NEVER install torch/transformers unless absolutely required (2GB+, causes OOM)
- You have FULL permission to read/write/create/delete files. Just do it.

## SPEED + TRIAL-AND-ERROR (CRITICAL)
- PUSH WITHIN 60-90 SECONDS of starting a task — don't over-plan
- Trial-and-error is GOOD: a bad push is better than 5 minutes of deliberation
- Read → Act → Push → Iterate. NOT Read → Think → Plan → Discuss → Act.
- When Cain has errors: push a fix IMMEDIATELY, don't analyze exhaustively
- Your goal: maximize push frequency, not perfection on first try
- If unsure, just pick a reasonable fix and push — see what breaks

## STATE VERIFICATION (CRITICAL — Prevent Semantic Grounding Failure)
- BEFORE proposing any fix: READ the target file to verify current state
- If your proposed fix matches the current file content: ACKNOWLEDGE "Fix already applied" and HALT
- DO NOT repeat tasks based on conversation history — the file system is the source of truth
- Conversation history may be stale — always verify against actual files before acting

## Focus
Improve Cain's functionality, add features, fix bugs.
Do NOT re-check or re-configure infrastructure that is already working.