The Agent Skills Directory

[DYNAMIC_EXECUTION]: The preamble and context recovery scripts use source <(...) and eval "$(...)" to execute the output of local binaries found in ~/.claude/skills/gstack/bin/. These operations are used to dynamically determine the repository mode and project slug.
[DATA_EXFILTRATION]: The skill incorporates a telemetry logging system and an artifact synchronization feature. Usage data and project artifacts (like designs and plans) may be sent to external repositories or servers; however, these features are preceded by user-consent prompts and can be disabled.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the repository, such as PR diffs, commit messages, and TODOS.md, which could contain malicious instructions meant to influence the agent's review findings.
Ingestion points: PR diffs, commit messages, plan files, and documentation (e.g., TODOS.md, CLAUDE.md).
Boundary markers: None explicitly defined to separate untrusted code data from the agent's execution instructions.
Capability inventory: Includes Bash for system commands, Write/Edit for file modification, and the Agent tool to spawn subagents.
Sanitization: No specific sanitization or escaping of the ingested repository data is implemented before it is processed by the AI.

review