mode-router

mode-router

A UserPromptSubmit hook (hooks/route.js) fires on every prompt. In auto mode it makes the model classify the request and invoke exactly one skill: coding task → ponytail (minimal code), everything else → caveman (terse output). The hook only routes; the two skills own their behavior. This skill reads and flips the control file that picks the mode.

Control file

$XDG_CONFIG_HOME/mode-router/state.json (or ~/.config/mode-router/state.json):

{ "mode": "auto" }

mode is one of: auto (default — model routes per request), caveman (force terse everywhere, regardless of request type), ponytail (force minimal-code everywhere), off (inject nothing). Missing or invalid → auto.

In auto, the hook stays silent on slash-command prompts — the user already dispatched to a skill, so classification never overrides it. A forced mode (caveman/ponytail) is a standing choice and still applies on slash prompts.

Spec-driven workflows

Auto routing is prompt-based, so it does not reach into a multi-turn spec-driven workflow (openspec, bmad, …) launched from a single slash command. Supported path: force the mode first — ponytail for the coding phase, caveman for analysis — then reset to auto. For direct prompts, the natural /clear between analysis and coding already re-classifies each phase.

Operations

1. Status — read the control file (report auto if absent) and state the active mode plus the routing rule above. Done when the user knows which mode is in force.
2. Set mode — write { "mode": "<value>" } to the control file, creating the directory if needed. Reject any value outside the four above. Done when the file holds the requested value.

Changes take effect on the next prompt — the hook re-reads the file every turn. Nothing is retroactive to the current turn.