The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the Bun installation script from the official domain (https://bun.sh/install). This is a well-known service associated with the project's runtime environment.
[REMOTE_CODE_EXECUTION]: Executes the downloaded Bun installer using bash. The skill includes a hardcoded SHA-256 checksum to verify the script's integrity before execution, reducing the risk of tampering during the download process.
[DYNAMIC_EXECUTION]: Employs eval to execute shell commands generated at runtime by local binaries such as gstack-slug. This allows the skill to load environment configurations dynamically from script outputs not defined in the plain text instructions.
[COMMAND_EXECUTION]: Uses process substitution to source shell environments from internal binaries like gstack-repo-mode to configure the repository state.
[DATA_EXFILTRATION]: Collects and transmits telemetry data, including skill name, duration, and session ID, to a remote endpoint. This behavior is gated behind an explicit user opt-in/opt-out prompt during the first run, allowing the user to manage metadata sharing preferences.
[PROMPT_INJECTION]: Provides detailed instructions that override default agent behaviors for the AskUserQuestion tool, including custom formatting requirements, plan-mode safety overrides, and fallback procedures for tool resolution.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Processes data from learnings.jsonl, CLAUDE.md, and project-specific test plans found in the ~/.gstack/projects/ directory.
Boundary markers: The skill does not define clear delimiters or specific instructions for the agent to ignore embedded commands within the ingested data.
Capability inventory: The skill has access to the Bash and Write tools, which could be used to execute commands or modify files if malicious instructions are present in the processed data.
Sanitization: No sanitization or validation of the external content is performed before it is used to influence the agent's actions.

qa-only