The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the amq command-line utility to coordinate tasks and messages between agents.
Evidence: Multiple instances of amq send, amq watch, amq thread, and amq drain are used throughout SKILL.md and references/spec-workflow.md to manage the design lifecycle.
Verification: The skill includes logic to verify the tool's presence using which amq and initializes the environment with amq coop init if necessary.
[PROMPT_INJECTION]: The workflow involves processing input from other agents, creating an indirect prompt injection surface.
Ingestion points: Data from partner agents is read into context using amq thread --include-body and amq drain --include-body (documented in references/spec-workflow.md).
Boundary markers: While explicit technical delimiters are not specified in the protocol, the skill instructs agents to perform and submit their own research before reading partner input to preserve perspective.
Capability inventory: The skill has access to codebase exploration tools for research and communication tools for agent-to-agent data transfer.
Sanitization: Security is managed via a mandatory human-in-the-loop gate; agents are explicitly forbidden from implementing designs based solely on agent-to-agent alignment.
[SAFE]: The design of the skill prioritizes user control and verification.
Structural Gates: The protocol requires agents to raise a to:user gate and wait for explicit human approval on a dedicated thread (gate/<topic>) before execution.
Role Separation: The 'initiator' agent maintains the user relationship, while 'partner' agents are restricted to research and review, preventing autonomous unauthorized implementation.

amq-spec