The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and act upon outputs from other AI agents (e.g., results from ak run and PLAN.md files), which constitutes a potential vulnerability surface.\n
Ingestion points: Agent execution results (result_text), execution transcripts, and shared planning documents (PLAN.md) are ingested across the orchestration loop.\n
Boundary markers: Instructions explicitly recommend using single quotes or quoted heredocs (e.g., <<'EOF') for task inputs to prevent shell expansion of agent-supplied text.\n
Capability inventory: The skill uses git for worktree management, ak for agent execution, and herdr for interactive pane management.\n
Sanitization: The instructions emphasize that agent outputs are 'claims' rather than evidence, require verification via repository test suites/linters before merging, and state that agent replies never grant permissions.\n- [COMMAND_EXECUTION]: The skill utilizes several command-line tools to perform its orchestration functions.\n
Evidence: Shell commands such as git worktree, ak pick, ak usage, and herdr pane are used throughout the documentation and scripts.\n
Safety Controls: The skill explicitly forbids the agent from answering credential prompts, creating profiles, or overriding user-configured permission postures (e.g., AGENTKIT_PERMISSIONS=ask).

agentkit-orchestrate