The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses eval on the output of the local script ~/.claude/skills/gstack/bin/gstack-slug. If this script generates output based on repository metadata or directory names, it could be exploited to execute arbitrary shell commands via command injection.
[COMMAND_EXECUTION]: User-provided arguments are passed directly to the execution of ~/.claude/skills/gstack/bin/gstack-gbrain-sync.ts via bun run. This creates an argument injection surface if the receiving orchestrator script does not implement strict input validation.
[COMMAND_EXECUTION]: The skill performs numerous executions of local binaries and scripts located in ~/.claude/skills/gstack/bin/, including gstack-config, gstack-team-init, gstack-gbrain-detect, and gstack-gbrain-repo-policy.
[DATA_EXFILTRATION]: The skill accesses and reads/writes to configuration directories in the user's home folder, including ~/.claude/, ~/.gstack/, and ~/.gbrain/. This involves reading state files and configuration data necessary for the tool's operation.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill reads the source_id from ~/.gstack/.gbrain-sync-state.json and queries the gbrain sources list for page_count values.
Boundary markers: It uses HTML comment delimiters () to manage the search guidance block in CLAUDE.md.
Capability inventory: The skill has broad capabilities including shell command execution (Bash), file modification (Write, Edit), and repository management (git).
Sanitization: No explicit sanitization or validation is performed on the data extracted from the JSON state files or command outputs before it is used in conditional logic or written to the CLAUDE.md file.

sync-gbrain