The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to manage code changes and repository state.
Evidence includes the use of git for branching, committing, and pushing changes to the alvarovillalbaa/agent-suite repository.
Evidence includes the use of the GitHub CLI (gh) to programmatically create Pull Requests with summaries of improvements.
The eval loop and autoresearch sub-flows execute scripts and testing tools (e.g., pytest, npm run build, python scripts/run_experiment.py) to measure and optimize file metrics.
[COMMAND_EXECUTION]: The skill documentation describes and encourages the use of the !command syntax within SKILL.md files to inject dynamic shell output into the agent's context at load time.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection (Category 8) due to its core function of autonomously mutating its own instructions and memory files based on external data.
Ingestion points: Data enters the system via 'discovery passes' of conversation history, user directions, and repository documentation files.
Boundary markers: The skill lacks explicit boundary markers or 'ignore' instructions when interpolating gathered evidence into the prompts used for generating file mutations.
Capability inventory: The skill possesses significant capabilities, including filesystem write access, shell command execution, and the ability to push code to remote repositories.
Sanitization: There is no evidence of sanitization or strict validation of the ingested evidence before it is used to propose and apply changes to the agent's logic or memory.
[EXTERNAL_DOWNLOADS]: The evaluation dashboard generated by the skill fetches external resources for visualization.
Evidence: The dashboard.html template loads the Chart.js library from a well-known CDN (jsdelivr).

auto-improve