The Agent Skills Directory

[PROMPT_INJECTION]: The SKILL.md file contains instructions that explicitly direct the agent to suppress its analytical and reasoning protocols. By framing the task as "mechanical work" and instructing the agent to "don't analyze each modified file" and "don't reason through a multi-step manual reproduction," the skill steers the AI away from its standard safety and evaluation behaviors.
[COMMAND_EXECUTION]: The skill utilizes spawnSync to execute a range of shell commands including git, node, gh, pnpm, and pgrep across several scripts (lib/cascade-template.mts, lib/cascade-tool-pins.mts, and lib/reconcile-lockfiles.mts). These commands are used to manage project worktrees, perform commits, and interact with repository tooling.
[EXTERNAL_DOWNLOADS]: The skill performs network-enabled operations such as git fetch and git push, and uses the GitHub CLI (gh) to create pull requests and monitor CI runs. These activities are directed toward the vendor's own GitHub repositories.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from various files and command outputs, creating a surface for potential indirect influence.
Ingestion points: Reads repository names from lib/fleet-repos.txt, tool versions from external-tools.json, and repository metadata from git status output.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are used when interpolating ingested data into shell command arguments.
Capability inventory: The skill possesses capabilities for file system modification, shell command execution, and network access via git and the GitHub CLI.
Sanitization: The skill uses standard path joining and string splitting for processing data but lacks explicit validation or escaping of project names or version strings before they are utilized in shell commands.

cascading-fleet