The Agent Skills Directory

[SAFE]: The skill implements a robust state management system using the Exarchos MCP suite, which centralizes workflow tracking and prevents fragmented or manual state management. This is consistent with the skill's primary purpose of saving and mutating workflow state.
[PROMPT_INJECTION]: The skill manages state that incorporates data from external sources such as pull request feedback, creating a surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through the updates parameter in the exarchos_workflow tool, specifically targetting fields like synthesis.prFeedback and tasks[].result as described in SKILL.md.
Boundary markers: The skill documentation does not currently specify the use of delimiters or 'ignore' instructions for data processed into the workflow state.
Capability inventory: The skill maintains control over workflow phase transitions, task status, and artifact locations, which can influence the behavior of the current agent or subagents.
Sanitization: Phase transitions are validated by a Hierarchical State Machine (HSM) to ensure prerequisites are met, though general field updates rely on the validation of the underlying MCP tool.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing optional companion tools via a remote command execution pattern.
Evidence: The mcp-tool-reference.md file suggests running npx create-exarchos to set up the environment, which involves fetching and executing code from the npm registry.

checkpoint