The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands and local binaries (e.g., gstack-slug, gstack-paths, gstack-config) to manage project state, configure routing, and resolve storage paths.
[COMMAND_EXECUTION]: Employs the eval command on the output of local binaries (e.g., eval "$(gstack-slug)") to dynamically set environment variables. While a common pattern in developer tools, this represents dynamic code execution within the agent's shell session.
[COMMAND_EXECUTION]: Includes instructions to modify repository files (.gitignore, CLAUDE.md) and perform Git operations (git rm, git add, git commit) as part of its migration and setup flows.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by ingesting untrusted data from the repository's history into the agent's context for summarization.
Ingestion points: Captures output from git log --oneline, git status, and git diff to provide context for the saved session.
Boundary markers: Uses simple text headers (e.g., === BRANCH ===, === STATUS ===) to delimit data fields, which may not prevent instructions embedded in commit messages or diffs from being interpreted by the model.
Capability inventory: The skill has the ability to execute shell commands, perform filesystem writes in the user's home directory (~/.gstack/), and modify the local Git repository.
Sanitization: Features a robust bash-side sanitization routine for user-supplied titles (using tr and cut allow-listing) to prevent command injection vulnerabilities during filename generation.

context-save