The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's preamble and various workflow steps execute the output of local binaries using eval and source <(...). This is primarily used to dynamically set environment variables and configuration based on the project context.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it processes untrusted data from multiple sources.
Ingestion points: It reads git logs, branch diffs, TODOS.md, plan files, and external review comments fetched via the Greptile API.
Boundary markers: The skill mitigates risks by dispatching subagents for specific tasks like coverage auditing and comment triage, providing them with strict instructions to "report only" and avoid direct code execution on the analyzed data.
Capability inventory: The skill possesses extensive capabilities, including full shell access (Bash), file modification (Write), and the ability to spawn subagents (Agent).
Sanitization: While it uses subagent isolation, there is no explicit evidence of sanitization for external content before it is incorporated into the LLM prompts.
[EXTERNAL_DOWNLOADS]: During the "Test Framework Bootstrap" phase, the skill can install external dependencies using standard package managers (e.g., npm, pip, gem, bun) depending on the detected project runtime.
[COMMAND_EXECUTION]: The skill performs numerous high-privilege shell operations, including fetching from remotes, merging branches, and pushing code to the repository's origin. These actions are fundamental to the skill's purpose as a deployment automation tool.
[DATA_EXFILTRATION]: The skill includes a telemetry system that logs usage metadata (such as skill name, duration, and session identifiers) to a remote server. This behavior is disclosed to the user with an explicit opt-in/opt-out prompt during the first run.

ship