The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Downloads and executes the Bun installation script from the official bun.sh domain. The installation process includes a mandatory SHA256 checksum validation to verify the script's integrity before execution.
[COMMAND_EXECUTION]: Employs eval and source commands to process the output of internal helper binaries (gstack-slug and gstack-repo-mode) located within the skill's local directory for environment setup.
[EXTERNAL_DOWNLOADS]: Fetches components from bun.sh and provides setup guidance for ngrok.com, both of which are established and well-known technology service providers.
[DATA_EXFILTRATION]: Records session metadata, skill usage statistics, and local repository names to analytics files on the local filesystem. Any transmission of this data to the vendor's servers is subject to a user-facing opt-in process.
[PROMPT_INJECTION]: Ingests project-specific data from CLAUDE.md and learnings.jsonl files. These files represent a surface for indirect prompt injection if they contain instructions authored by an untrusted source.
Ingestion points: CLAUDE.md, learnings.jsonl (processed in the SKILL.md preamble)
Boundary markers: Not present for context ingestion
Capability inventory: Shell access (Bash), file reading (Read), interactive questioning (AskUserQuestion)
Sanitization: No specific filtering or escaping logic detected for ingested context files

pair-agent