The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a multi-agent workflow where a coordinator and sub-agents ingest task descriptions and research findings, creating a potential surface for indirect prompt injection.
Ingestion points: The agent processes user-provided task descriptions and reads from files like findings.md during the discovery and execution stages.
Boundary markers: The instructions suggest using "bounded tasks" in Codex profiles, but the logic does not enforce strict structural delimiters or escaping for interpolated external data.
Capability inventory: The skill is authorized to spawn sub-agents (runtime:spawn-agents) and perform filesystem writes (filesystem:write-planning-artifacts), which could be misused if malicious instructions are ingested.
Sanitization: There are no explicit sanitization or filtering mechanisms defined in the prompt instructions to handle potentially adversarial content in the input tasks.
[COMMAND_EXECUTION]: The project includes a shell installation script and a Python test suite that executes local commands.
Evidence: scripts/install.sh performs directory management, file copying, and validation using standard POSIX utilities.
Evidence: tests/test_installer.py utilizes subprocess.run to verify the installation script's behavior within isolated temporary directories.
Note: These actions are standard for software distribution and automated testing and do not show signs of malicious intent.
[EXTERNAL_DOWNLOADS]: The installation utility allows for fetching skill components from a remote source.
Evidence: scripts/install.sh can download SKILL.md and README.md from the author's public GitHub repository at https://raw.githubusercontent.com/KimYx0207/Kim_Service/main/ when the --from-github flag is used.

agent-teams-playbook