The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a multi-agent coordination protocol that relies on reading and responding to messages from other agents. This creates a surface where one agent could influence another through malicious instructions embedded in messages.
Ingestion points: Found in SKILL.md in the 'Receive a message' section, where the agent uses herdr agent read to ingest transcripts from other panes.
Boundary markers: The skill defines a formal envelope [herdr peer ...] to distinguish peer messages from other content.
Capability inventory: The agent can execute shell commands via the herdr CLI, including agent prompt, agent read, and agent send-keys across different panes.
Sanitization: The instructions include a requirement to manually inspect the origin pane's context and transcript if a message is suspicious, and to consult the user for sensitive or destructive actions.
[COMMAND_EXECUTION]: The skill uses the herdr CLI to interact with the environment and other agents. It constructs shell commands using variables like <session>, <origin-pane>, and <target-pane> derived from the coordination protocol. While these are intended for legitimate orchestration, they represent a surface for command injection if input parameters are not strictly validated by the underlying platform.

herdr-cross-agent-coordination