The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to manage and execute shell commands via tmux or cmux to control worker agent sessions. This capability is central to the skill's purpose as an orchestrator but grants the agent broad control over local processes.
[DYNAMIC_EXECUTION]: The orchestrate.sh script dynamically sources agent profile files from a computed path in the user's state directory ($ORCH_HOME/profiles/<name>.profile). This allows for extending the tool with new agent definitions but involves the execution of shell code from a local directory.
[INDIRECT_PROMPT_INJECTION]: The skill implements a supervision pattern where a 'control tower' agent reads and interprets the transcript of 'worker' agents, creating a surface for indirect prompt injection if a worker agent processes malicious data.
Ingestion points: Worker agent output is captured via terminal screen scraping (tmux capture-pane or cmux read-screen) and presented to the supervisor agent through subcommands such as poll, logs, and run.
Boundary markers: The instructions specify a sentinel string (>>> NEEDS_HUMAN:) to signal when a worker requires intervention, but they do not provide formal delimiters (e.g., XML tags) to separate worker logs from the supervisor's instructions.
Capability inventory: The supervisor agent can spawn new processes, send keyboard inputs to existing sessions, and manage lifecycle operations for all workers via the orchestrate.sh wrapper.
Sanitization: The orchestrate.sh script includes a strip_ansi function that cleans terminal escape codes from the captured output, which provides basic sanitization before the data is ingested by the agent.

orchestrate