How caoshiyi/eat works
EAT is a Claude Code plugin for running evolving teams of coding agents against a user’s task. The documented loop is: Claude Code helps prepare the task and evaluation, launches a run, and a supervisor diagnoses stalled iterations from transcripts and rebuilds the team. The tree also includes clean-room isolation and run-verification/curve utilities. The README supports the main workflow; sampled source supports the proxy and utility behavior. Launch-script internals and other unsampled wiring are not asserted.
Main entry points: SUPERVISOR.md Supervisor · SKILL.md EAT skill · allowlist_proxy.py Allowlist proxy · tcp2unix.py TCP-to-Unix bridge
Components 18
Agent orchestration
- EAT skill
SKILL.md - Supervisor
agent_team/SUPERVISOR.md - Team config parser
agent_team/parse_team_config.py - Team launcher
agent_team/launch_team.sh - Stateless launcher
agent_team/launch_stateless.sh - Result extractor
agent_team/extract_result.py
Clean-room controls
- Run sandbox
cleanroom/sandbox.sh - Allowlist proxy
cleanroom/allowlist_proxy.py - TCP-to-Unix bridge
cleanroom/tcp2unix.py - Trace audit
cleanroom/audit_clean_traces.py - Code overlap audit
cleanroom/audit_code_overlap.py
Run operations
- Task fetcher
ops/fetch_task.sh - GPU runner
ops/gpu_run - Iteration verifier
ops/verify_iter.py - Curve exporter
ops/curve_data.py
Other components
- User
- Claude Code
- Anthropic / Claude API
Connections 7
- User → Claude Code: requests run
README.md - Claude Code → EAT skill: invokes skill
README.md - EAT skill → Supervisor: starts workflow
README.md - Supervisor → Team launcher: launches team
README.md - Supervisor → Supervisor: diagnoses and rebuilds
README.md - Allowlist proxy → Anthropic / Claude API: forwards allowed traffic
cleanroom/allowlist_proxy.py - TCP-to-Unix bridge → Allowlist proxy: bridges socket
cleanroom/allowlist_proxy.py