The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses eval and source to execute environment configurations and path logic generated by its own internal binaries, such as gstack-slug and gstack-paths.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local repository environment to generate context summaries.
Ingestion points: Reads project state from git status, git diff, git log, and CLAUDE.md.
Boundary markers: Uses Markdown headers and YAML frontmatter to delimit saved context, providing structural separation.
Capability inventory: Performs file writes to ~/.gstack/ and the project root, executes git commands, and runs internal binaries.
Sanitization: Implements title sanitization using shell utilities like tr and cut to prevent injection via user-supplied session names.
[COMMAND_EXECUTION]: The skill executes various local shell commands and binaries to manage session state, clean up temporary files, and handle configuration updates.
[DATA_EXFILTRATION]: The skill includes opt-in telemetry and artifact synchronization. Telemetry data is sent via a local binary to the vendor's infrastructure, while artifact sync targets GitHub, a well-known and trusted service. Both features require explicit user consent via interaction.
[PERSISTENCE]: With user approval, the skill can modify the project's CLAUDE.md file to integrate skill routing rules and update .gitignore to manage its own internal directory.
[EXTERNAL_DOWNLOADS]: The skill references and fetches information from external sources, including a vendor-associated domain for documentation and GitHub for configuration and updates.

context-save