The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses dynamic execution patterns to run code generated by local binaries at runtime.\n
Evidence: the preamble uses eval \"$(~/.claude/skills/gstack/bin/gstack-slug)\" and source <(~/.claude/skills/gstack/bin/gstack-repo-mode) to execute shell commands generated by binary outputs.\n- [COMMAND_EXECUTION]: The skill invokes numerous custom binaries for operational and configuration tasks.\n
Evidence: calls to gstack-update-check, gstack-config, gstack-session-kind, gstack-telemetry-log, gstack-slug, gstack-learnings-search, and gstack-brain-sync located in the vendor's local binary directory ~/.claude/skills/gstack/bin/.\n- [DATA_EXFILTRATION]: The skill manages features that transmit project and usage data externally.\n
Evidence: 'Artifacts Sync' feature publishes user artifacts (plans, designs) to a private GitHub repository; the skill also includes an opt-in telemetry system sending metrics to a remote endpoint via the gstack-telemetry-log binary.\n- [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface due to the ingestion of untrusted external and local data during debugging.\n
Ingestion points: Reads local project learning files (learnings.jsonl), timeline logs, and output from the WebSearch tool.\n
Boundary markers: No explicit delimiters are specified for separating untrusted data from instructions.\n
Capability inventory: The skill has access to shell execution, file system edits, and network search tools.\n
Sanitization: Sanitization is requested for web search queries but is not enforced for data ingested from the filesystem.\n- [METADATA_POISONING]: The skill automatically modifies project metadata to influence future agent behavior.\n
Evidence: Offers to append 'Skill routing' rules to the CLAUDE.md file and commits the change, which directs how the agent selects skills for various request types in that project.

investigate