The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and follow instructions from a workspace file named custom.md specifically during the 'Settlement' phase. This creates a vulnerability where untrusted content in the project directory can influence or override the agent's behavior during high-stakes actions like bounty processing and code PR creation.
Ingestion points: custom.md, CLAUDE-activeContext.md, MEMORY.md (via /sync).
Boundary markers: None. The skill is explicitly told to "follow instructions" from the file.
Capability inventory: Subprocess execution (git, gh, rm), tool-based financial operations (contract_submit, contract_settle), and sub-agent spawning (Task).
Sanitization: None detected.
[COMMAND_EXECUTION]: The skill performs several shell operations including pushing code to remote repositories (git push), creating pull requests (gh pr create), and deleting specific configuration files in the user's home directory (rm -f ~/.claude/ralph-loop.local.md). While these appear consistent with the skill's stated purpose, the file deletion occurs in a hidden system-level directory.
[DYNAMIC_EXECUTION]: The skill programmatically spawns a secondary 'finality-agent' via the Task capability. While used here for verification, the ability to spawn autonomous sub-agents with dynamically generated prompts increases the complexity and risk profile of the skill's execution.

stop