The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill maintains an attack surface for indirect prompt injection by ingesting and acting upon data from the local file system and user input.
Ingestion points: The orchestrator reads progress.txt and project metadata files (such as package.json, requirements.txt, and Gemfile) to detect project context and determine the current development phase (described in SKILL.md and references/pipeline/entry-point-detection.md).
Boundary markers: The instructions do not define strict delimiters or "ignore embedded instructions" markers when reading the contents of progress.txt or external dependency files.
Capability inventory: The skill is capable of reading and writing local files (specifically for state management) and invoking a wide range of phase-specific skills (e.g., vibe-coding-build, vibe-coding-deploy-vercel).
Sanitization: There is no explicit requirement for the agent to sanitize or validate the content of the project files before using that data to decide its routing or next automated steps.

vibe-coding-orchestrator