The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's core functionality relies on executing the herdr and orca CLI tools to manage agent sessions, split panes, and communicate across machines.
Evidence: Found in scripts/herdr-resolve.py, scripts/herdr-send.py, and scripts/herdr-peer.sh, which invoke subcommands like agent list, pane split, and machine list via subprocess.run or shell execution.
[INDIRECT_PROMPT_INJECTION]: By establishing a protocol for inter-agent communication, the skill creates an attack surface where a session could be influenced by malicious data or instructions sent from another session.
Ingestion points: Data enters the agent's context through messages received via the herdr agent prompt or orca terminal send mechanisms described in SKILL.md and references/orca-routing.md.
Boundary markers: The protocol defined in SKILL.md includes an 'anti-hallucination' requirement for agents to verify prompt data against the repository and code sources.
Capability inventory: The skill environment allows for shell command execution (Bash), file reading (Read), and environment access (Env), which could be leveraged if an injection succeeds.
Sanitization: scripts/herdr-send.py implements quoting strategies using temporary files and shell variables to ensure messages are delivered without causing command argument injection.
[DYNAMIC_EXECUTION]: Python scripts in the scripts/ directory construct and execute shell commands dynamically based on session IDs, pane IDs, and machine labels provided at runtime.
Evidence: scripts/herdr-resolve.py and scripts/herdr-send.py use subprocess.run to call herdr with arguments determined by user or agent input.

multi-agent-collab