The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on external command-line utilities (aqe and ruflo) for critical operations like memory retrieval and state persistence.\n
Evidence: steps/01-flag-detection.md executes aqe memory search and npx --no-install ruflo memory search.\n
Evidence: steps/07-learning-persistence.md executes aqe memory store and aqe memory share.\n- [DYNAMIC_EXECUTION]: Shell command templates within the skill use string interpolation to inject variables, which is a known pattern for command injection vulnerabilities if the variables contain user-controlled content.\n
Evidence: steps/07-learning-persistence.md constructs a shell command using --key \"qcsd-refinement-${storyId}-${Date.now()}\" and a JSON --value string, both of which are susceptible to shell escaping and injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted STORY_CONTENT to gate its execution logic via flags (e.g., HAS_API), creating a risk that instructions inside the data could manipulate the swarm's behavior.\n
Ingestion points: The STORY_CONTENT provided by the user is processed in steps/01-flag-detection.md.\n
Boundary markers: No delimiters or specific instructions are provided to the agent to prevent it from obeying instructions embedded within the user story text.\n
Capability inventory: The skill has access to shell execution tools (aqe, ruflo), can write multiple files to the filesystem (OUTPUT_FOLDER), and can spawn various sub-agents with differing capabilities.\n
Sanitization: There is no evidence of input sanitization or validation performed on the story content before it is used for logic branching or stored in persistence layers.\n- [EXTERNAL_DOWNLOADS]: The skill introduces a dependency on the Node.js package ruflo via npx.\n
Evidence: steps/01-flag-detection.md references ruflo, which is an uncommon/unverified package. While usage via --no-install mitigates immediate remote downloads, it remains an unverifiable dependency.

qcsd-refinement-swarm