The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and analysis of external, unmanaged project source files, creating a potential surface for indirect prompt injection.
Ingestion points: Project files (intent prose, code, tests, configuration) are read and cataloged into a sources.md manifest within the .devflow/ directory.
Boundary markers: The skill mandates the use of specific classification units (observed fact, interpretation, unknown, conflict) to process ingested data, as defined in SKILL.md and references/project-knowledge.md.
Capability inventory: The agent is authorized to read project source files, write documentation to the .devflow/ directory, and interact with Git metadata.
Sanitization: The instructions in references/communication.md and references/product-document.md emphasize semantic reconstruction over verbatim copying, which serves as a natural filter against embedded instructions.- [DYNAMIC_EXECUTION]: The skill utilizes local JavaScript scripts for runtime integrity checks and command execution within the project environment.
Evidence: scripts/runtime/common.mjs defines a runChild function using node:child_process.spawn. scripts/run.mjs is the entry point for performing artifact integrity verification.
Context: These utilities are part of the skill's framework for ensuring file consistency. The code includes security best practices, such as path normalization and realpath resolution in resolveProjectPath to mitigate path traversal vulnerabilities.

adopt