The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a workflow that ingests codebase content (untrusted data) to generate implementation plans, creating a surface for indirect prompt injection.
Ingestion points: Research phase instructions in SKILL.md and references/architecture.md direct the agent to read files, search symbols, and inspect diagnostics.
Boundary markers: The skill enforces a strict planning-only boundary that forbids file mutation (except for plan artifacts) and requires explicit authorization transitions before implementation.
Capability inventory: The agent can write Markdown plan files, execute shell commands in the execution phase, and delegate tasks to subagents.
Sanitization: Artifact structure is validated by scripts/plan_artifact.py, although the content itself is derived from the repository environment.
[DYNAMIC_EXECUTION]: The skill dynamically resolves the path to its utility script plan_artifact.py and executes it.
Evidence: SKILL.md contains logic to locate the helper script in the workspace or agent configuration directories before running it with python3.
[COMMAND_EXECUTION]: The skill uses local shell commands for environment discovery and artifact management.
Evidence: Execution of git rev-parse to find the workspace root and python3 for artifact initialization and checking.

plan-mode