The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository source code through analysis tools, which constitutes an indirect prompt injection surface.
Ingestion points: The agent ingests project source files, semantic symbol maps, and configuration files like sgconfig.yml and codegraph.json (documented in SKILL.md and references/usage-playbook.md).
Boundary markers: The instructions explicitly mandate the redaction of secrets, static headers, and customer data from persisted or public artifacts (documented in Safety rules).
Capability inventory: The skill utilizes shell command execution for tool installation (npm, cargo, pipx), writes configuration files to the project root (e.g., .mcp.json, config.toml), and performs repository-scale indexing (documented in references/setup-and-mcp-config.md).
Sanitization: The instructions include protocols for redacting credentials and using bounded search patterns to limit data exposure.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch binaries and packages from external registries and repositories.
Tool Provenance: Fetches official releases from github.com/colbymchenry/codegraph and uses public registries (npm, PyPI, Cargo) for tool installation.
Verification: Recommends manual checksum verification (SHA256) for standalone release assets to ensure integrity before extraction.
[COMMAND_EXECUTION]: The skill provides extensive shell command templates for setup, diagnostics, and structural code searching.
Safe Execution: Commands are designed to use project-local shims (e.g., node_modules/.bin/codegraph) to prevent shadowing by untrusted binaries on the PATH.
Behavior Control: Routinely uses environment variables like CODEGRAPH_TELEMETRY=0 to suppress unwanted data transmission during automated checks.

codegraph-ast-grep