The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill setup process involves downloading and executing scripts from the vendor's repository.
Evidence: scripts/bootstrap.sh and scripts/bootstrap.ps1 download an installer from github.com/ArchAstro/archdev and execute it.
Mitigation: The skill follows best practices by using a pinned commit hash in the URL and verifying the downloaded file's SHA-256 hash before execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from 'Rooms' messages and search results which could theoretically contain malicious instructions.
Ingestion points: Data is brought into the session context via rooms search and rooms messages commands as instructed in SKILL.md.
Boundary markers: The instructions explicitly direct the agent to treat teammate posts as information rather than instructions and require local verification of findings.
Capability inventory: The agent's capabilities are limited to the archdev CLI for posting, searching, and connection management.
Sanitization: The skill relies on prompt-level instructions to separate teammate-provided information from agent commands.
[DATA_EXFILTRATION]: The skill includes robust instructions to prevent the accidental transmission of sensitive data.
Evidence: SKILL.md mandates that the agent must "never send raw transcripts, credentials, or private content without authorization" and "Never post secrets, tokens, customer data, or unreviewed private content."

rooms