/run-agent

Normally this skill never reaches a model: the UserPromptExpansion hook in settings.json (matcher run-agent) runs run-agent.sh --relaunch on the typed words itself and shows its output. You are reading this only because the hook did not run (older settings, hook disabled): then run exactly one Bash call and then stop. Do not parse, reason about or confirm anything yourself: a clean parse runs, schedules or opens its window at once; the script hands the questions it cannot answer (confirm typed, an unplaceable word, no agent named, a folder to create, which job to cancel) to a short haiku session (claude --model haiku --effort low) that asks, then runs or schedules.

~/.claude/skills/run-agent/scripts/run-agent.sh --relaunch '<raw $ARGUMENTS>'

Single-quote the arguments (escape ' as '\'', keep newlines). Relay its output as is. Exit 0 means this pane now runs the agent (or the confirm step when there was a question, or a new tmux window opened) or the script answered itself (jobs, cancel, dry, scheduled #N, errors); exit 1 = not inside tmux, relay the run-agent ... line to type by hand; exit 2 = relay the message.

Forms: <agent> ["prompt"] [in <dir>] [split|window] [model/effort words] [at <time> | in <delay> | after window [until <time>]] [dry] [confirm], jobs, cancel [id]. split and window both open a brand-new tmux window and this session keeps going; without them the pane is replaced. A launch never adds a pane to an existing window. Runs joined by a lone + (planner "Plan A" in ~/a + planner "Plan B" in ~/b + ...) start together, at once, in one new window tiled (four = 2x2), each in its own folder. Everything starts without asking; confirm (or --confirm) anywhere outside the quotes asks first (a + group is asked about once). The old force/now/go/! words are no longer special: now means at once, the rest are asked about like any unplaced word.

Time words resolve in parse.py without a model: at 9pm, in 2h, in 2 hours 10 min, tomorrow [at] 1am, 1am tomorrow, tonight / this evening (7 PM), today at 5pm, noon, midnight, a weekday name (the next one), after window, N h|min before|after the window (or … the current window, … window ends/resets), N h before|after <clock>. jobs shows each run's time and runner state (sleeping (pid N), dead, errored: <last log line>, waiting: no pane). When the parser cannot place a time word the haiku confirm step reads it (it is told the current ET time and the window reset) and hands --step '<words> | time: at <clock>' back; it never asks the user when the window ends. The window reset comes from the usage API; when that fails (429, no network) the last live answer, then an estimate from session transcripts, then now + 5 h stand in, labelled, so a window-relative time always schedules. From an agent session (worker, evaluator, planner, helper) split, window and + are answered ignored: agents replace their own pane: an agent only ever replaces its own pane, now or at the scheduled time; if that pane is gone at fire time the run waits (jobs.log: waiting: no pane) until a human /run-agent or run-agent call on that tmux server opens it in a new window.

(Also: S --parse '<raw>' prints the parse as JSON; S --handoff <agent> ["prompt"] [-- flags] is the worker's usage-limit handoff; S --next <agent> ["prompt"] prints an agent's closing next step (short sentence, /run-agent line, Without run-agent: claude … line; no model/effort for worker and evaluator); bin/run-agent is the same skill as a shell command.)