The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses eval to process output from local binaries like gstack-slug and gstack-paths to set environment variables. While this pattern introduces dynamic execution, it is localized to the vendor's own tools.
[EXTERNAL_DOWNLOADS]: The skill mentions external dependencies and tools such as Aside and Codex. It provides clear instructions for the user to install these themselves rather than performing automated downloads.
[DATA_EXFILTRATION]: The skill sends data to external services for web research and model-based reviews. This is mitigated by explicit user consent prompts, sanitization of search queries, and the use of the gstack-redact tool to filter sensitive information.
[COMMAND_EXECUTION]: The skill executes multiple local binaries from the ~/.claude/skills/gstack/bin/ directory. These are part of the primary framework functionality for telemetry, caching, and logging.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection via web search results and external model outputs.
Ingestion points: Web search results (Aside/WebSearch), external model responses (Codex), and local repository files.
Boundary markers: Explicit instructions to ignore instructions in external data and specific file path exclusions in review prompts.
Capability inventory: Full shell execution (Bash), file write/edit, and network egress via tools.
Sanitization: Implements query sanitization and uses gstack-redact to prevent leaking secrets.

office-hours