The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a multi-agent environment where an agent reads and processes outputs from external sources and other agents, creating an attack surface for indirect instructions.
Ingestion points: Data enters the agent's context through tools like pasture pane tail (reading terminal output), herdr agent read (reading another agent's response), and pasture web read or the agent-browser skill (reading web content).
Boundary markers: There are no explicit instructions to use delimiters or specific safety wrappers when the agent processes output from these external sources.
Capability inventory: The skill possesses extensive capabilities including shell command execution (herdr, pasture-*, git, sl, bazel, uv, pnpm, bun), file system modifications within git worktrees, and the ability to trigger actions in other agents or web browsers.
Sanitization: The instructions do not define sanitization or filtering logic for data ingested from other panes or agents before it is interpreted by the LLM.
[COMMAND_EXECUTION]: The skill is heavily reliant on the execution of local CLI tools to manage the development workflow.
Evidence: It explicitly instructs the agent to use a variety of local scripts (pasture-task, wt, sl) and package managers.
Safety Measures: The skill includes a built-in safety protocol, advising the agent to use --dry-run for write operations and requiring explicit user approval (--yes flag) for commands labeled as 'external'.

pasture