The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently uses the Bash tool to execute local Python and shell scripts (e.g., idea_provenance.py, ar-workflow-engine.py, ar-preflight-mcp.sh, ar-gemini-monitor.py). These scripts are integral to the skill's function for initializing projects, managing a task queue, and monitoring experiment logs.
[PERSISTENCE]: The skill implements a persistence mechanism known as 'Ralph Loop' by writing configuration to .claude/ralph-loop.local.md and executing a setup script found in the user's plugin cache (~/.claude/plugins/). This allows the agent to be re-invoked automatically to continue long-running workflows. It also starts a background monitoring process (ar-gemini-monitor.py) that runs independently of the main agent session.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests user-provided idea_file content during Phase 0.
Boundary markers: The ingested idea text is propagated to sub-agents (planner, coder, reviewer) as part of their instructions. There are no explicit structural delimiters or warnings to ignore malicious instructions potentially embedded within the idea file.
Capability inventory: The coordinator has the ability to execute shell commands, read/write project files, and spawn sub-agents with code-writing and execution permissions.
Sanitization: While an inspection script (idea_provenance.py) validates the file's integrity and provenance, it does not perform semantic sanitization to strip or neutralize potential prompt injection attacks within the research ideas.

ar-coordinator