The Agent Skills Directory

[COMMAND_EXECUTION]: The preamble and main workflow perform extensive execution of local binaries belonging to the gstack toolset (e.g., gstack-update-check, gstack-config, gstack-slug, gstack-brain-sync).
[COMMAND_EXECUTION]: The skill uses eval and shell substitution source <(...) to execute shell code dynamically generated by the gstack-slug and gstack-repo-mode binaries in SKILL.md.
[DATA_EXFILTRATION]: The skill includes a telemetry system that sends usage data (skill name, duration, outcome, session ID) to a remote endpoint. While this is gated by a user prompt, the endpoint is not explicitly disclosed in the script.
[DATA_EXFILTRATION]: The 'GBrain Sync' feature is designed to publish session memory—which may include code snippets and technical decisions—to a private GitHub repository. This functionality is presented as an opt-in feature.
[PROMPT_INJECTION]: The skill contains strong directives to the agent to 'Treat the skill file as executable instructions, not reference' and to override standard behaviors based on configuration (e.g., EXPLAIN_LEVEL: terse).
[PROMPT_INJECTION]: Indirect prompt injection surface detected.
Ingestion points: Reads project-level files (CLAUDE.md), internal data logs (learnings.jsonl), and saved checkpoint files (checkpoints/*.md).
Boundary markers: Missing for several processed files, which could allow embedded instructions to influence the agent.
Capability inventory: The skill has broad capabilities including shell execution (Bash), file writing, and network operations.
Sanitization: Title inputs are sanitized via shell utilities (tr, cut) before being used in filenames.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to check for updates (gstack-update-check) and sync repository state via git (gstack-brain-sync).

context-save