The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md and adapters/claude-code.md files utilize the dynamic context injection syntax to execute a shell command at load time. This command executes a Python script (retrieve.py) using the ${CLAUDE_SESSION_ID:-} shell variable.
Evidence: !python3 ~/workspace/projects/faion-net/faion-network/skills/faion/scripts/retrieve.py "${CLAUDE_SESSION_ID:-}" in SKILL.md.
[COMMAND_EXECUTION]: The skill includes several complex administrative shell scripts located in hidden directories (.research, .reclass, .refactor) intended for repository maintenance. These scripts perform powerful operations including rm, mkdir -p, and git commands (git mv, git commit, git merge, git push).
Evidence: move.sh, build-queue.sh, heuristic.sh, and filter.sh scripts.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data such as user session transcripts and feature briefs, which are then used to inform tool use.
Ingestion points: Untrusted data enters the context via the retrieve.py script which reads session transcripts, and via user-provided briefs in methodology files.
Boundary markers: The methodology files (e.g., knowledge/ai-agents/agent-replay-harness-cookbook/content/03-failure-modes.xml) contain explicit instructions to the agent to treat inputs as DATA and ignore instruction-shaped text (e.g., 'ignore previous instructions').
Capability inventory: The skill frontmatter in SKILL.md explicitly allows shell execution via Bash(python3:*).
Sanitization: No programmatic sanitization or escaping of the ingested transcript data is demonstrated in the provided file contents.
[EXTERNAL_DOWNLOADS]: The documentation frequently encourages the installation of external Python packages for agent development.
Evidence: Recommendations in agent-patterns/methodology-v1.xml.bak to pip install packages like langchain-cli, litellm, and instructor.

faion