The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates an indirect prompt injection surface by dispatching mission prompts to sub-agents. If these prompts are sourced from untrusted external content, they could be used to manipulate the behavior of the agents spawned.
Ingestion points: The plan_project and create_mission tools accept a prompt string (SKILL.md).
Boundary markers: Absent; there are no instructions to use delimiters or ignore directives to separate the mission prompt from internal agent instructions.
Capability inventory: Agents dispatched via dispatch_mission operate in git worktrees with "full tooling," allowing for file system modification and tool execution (SKILL.md).
Sanitization: No evidence of sanitization or validation of the input prompt is present before it is sent to the agents.
[COMMAND_EXECUTION]: The documentation provides a command to configure a local MCP server via the Claude CLI (claude mcp add devfleet), which connects to a transport service on localhost (SKILL.md).

claude-devfleet