The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes shell commands via node:child_process.spawnSync to interact with the herdr CLI for window management and git for repository information. This is localized in scripts/lib/platform.mjs and used throughout the coordination logic.
[INDIRECT_PROMPT_INJECTION]: The orchestration logic involves reading Markdown reports and terminal screen content generated by sub-agents which are controlled by model providers. This creates an ingestion surface for potentially untrusted instructions from the model outputs.
Ingestion points: Report files are read in scripts/lib/commands/collect.mjs, and terminal screens are captured in scripts/lib/herdr.mjs via agentRead.
Boundary markers: The skill does not use explicit boundary delimiters or instructions to ignore embedded prompts when processing sub-agent output.
Capability inventory: The orchestrator can execute CLI tools (herdr, git) and write files to the project's scratch state directory.
Sanitization: The skill implements secret redaction (e.g., stripping strings matching API key patterns like sk- or Bearer) to prevent credential leakage during log recording or status updates.

herdr-agents