A template for an agentic Cyber Reasoning System (CRS): a Claude Code agent that reads a target's source, drives analysis tools (fuzzing, static analysis, coverage, a debugger), and produces verified crashing inputs (PoVs). It plugs into the oss-crs framework as a bug-finding CRS.
Start with the tutorial, which walks you through extending this template one capability at a time.
builder/: build-phase scripts, each producing one build variant of the target via OSS-Fuzzcompile.oss-crs/: the oss-crs integration manifest (crs.yaml,dockerfiles/, example compose).tutorial/: the 5-chapter hands-on tutorial for building this template into an effective CRS.crs/: the CRS implementation.CLAUDE.md.j2is the agent's standing-instructions template. Its subdirectories:crs/entrypoint/: the run-phase entrypoints plus theentrypoint.shmultiplexer (run_crs,run_crs_subagents,run_crs_langgraph,idle,draw_graph,test_*).crs/src/: runtime and prompt plumbing (runtime.py,prompts.py,claude_node.py).crs/tools/: thecrs-*CLI implementations (crs-fuzz,crs-afl,crs-codeql,crs-coverage).crs/skills/: Claude Code skill docs.crs/roles/: entrypoint role/system prompts.crs/agents/: Task-tool subagent definitions.