wochap/nix-config/skill

agents

Use to delegate a self-contained task to another coding agent session, e.g. work in another repo (-C <dir>) or a second opinion, without filling your context.

agents run -q "<prompt>"                 # stdout: final answer only; stderr: "session: <id>"
agents run --json "<prompt>"             # stdout: {"id","agent","model","result","costUsd","durationMs","status"}
agents run -q -C ~/repo -m claude-sonnet-5 -e high "<prompt>"
agents run -q -a pi "<prompt>"           # another agent (default: claude)
agents run -q -a agy "<prompt>"          # agy (default: gemini-3.8-flash)
echo "<long prompt>" | agents run -q -   # prompt from stdin
agents run -q -r <id> "<follow-up>"      # continue a session headless
agents run -q --no-tools "<question>"    # answer only, no tools (pass again on -r)
agents last <id>                         # last message of a session
agents ls                                # sessions of the last 7 days, --full for all

• Agents: claude (default), pi, agy (default model gemini-3.8-flash). -m defaults to the chosen agent's model.

-e (low, medium, high, xhigh, max) defaults to the agent's own setting. -r, last take the session's own agent; no -a needed.

• Ids may be a unique prefix.
• Always pass -q or --json; the prompt must be self-contained (the agent has none of your context).
• agents run blocks until the agent finishes; exit code 1 means the agent failed.
• Tell the user the session id: they can follow it live with agents watch <id>

or take it over with agents attach <id>.

• Never run agents attach or agents watch yourself; both are for the user's terminal.
• agents serve is a long-running OpenAI-compatible server for chat GUIs; never start it yourself.
• Taking over a running session (ctrl+t in the run view, ctrl+z to detach

back) needs a person at a TTY. You never press these keys; with -q or piped stdin there are none.