The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses the herdr CLI to manage agent lifecycles, terminal panes, and Git worktrees. Commands include agent start, agent prompt, pane split, and worktree create. The instructions mandate environment validation (HERDR_ENV=1) and identity verification before executing mutations.
[DYNAMIC_EXECUTION]: The skill executes a bundled Node.js script, scripts/publish-report.mjs, to validate and publish agent activity reports. The script follows security best practices by using O_NOFOLLOW to prevent symlink attacks and implementing path traversal checks for file lists.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to orchestrate multiple agents (Codex, Claude, AGY) and process their outputs, which creates an ingestion surface for potentially malicious instructions embedded in agent callbacks or terminal transcripts.
Ingestion points: Agent callbacks via herdr agent prompt, terminal history via herdr agent read, and structured producer reports in YAML/JSON format.
Boundary markers: The skill uses specific command prefixes (e.g., /herdr (/absolute/path/...)) and a structured 'authority block' to distinguish instructions from data.
Capability inventory: The skill can launch new interactive sessions, inject text/keys into terminal panes, and write report files to the filesystem.
Sanitization: While the publish-report.mjs script performs strict schema validation on report metadata, it does not sanitize the natural language content of summaries or evidence strings against embedded instructions.

herdr