The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute shell and Node.js scripts located within the repository's scripts/ directory, specifically scripts/link-skills.sh and node scripts/validate-skills.mjs. These executions are triggered during the 'Apply Cascade' phase of the workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from files in the .agents/feedback/ directory to determine which code changes to apply. This creates a surface for indirect prompt injection where malicious feedback could influence the agent's actions.
Ingestion points: Feedback files located in .agents/feedback/.
Boundary markers: The skill does not define specific delimiters for the feedback data, although it requires a manual approval step (Step 3: 'Présenter le plan') before any writes or executions occur.
Capability inventory: The skill has the capability to write to the file system (modifying SKILL.md, README.md, CHANGELOG.md, etc.), create new skill files, and execute shell/Node.js scripts.
Sanitization: There are no explicit instructions to sanitize or validate the content of the feedback files before processing.
[DYNAMIC_EXECUTION]: The skill invokes the Node.js test runner (node --test) on the repository's test suite after applying modifications. If the modifications were influenced by malicious feedback, this could lead to the execution of code derived from untrusted input during the testing phase.

improve-mabza