Steering
Read your own reins.
At your prompt, a live ๐ steer X% (taut ๐ชข past half your turns): how much you
redirected the agent this session. Watch it fall as the agent picks up your cues. The check guards
the turns you did not steer, so a miss there is the risk you can see.
Watch the rein go slack.
One line between you and the agent. It pulls taut when you steer hard and goes slack as the agent holds your cues. The sag is the steer rate falling.
๐ชข steer 88% ยท tight rein, early
๐ steer 12% ยท loose rein, the goal
Tune the judge for your tokens.
A small model scores each turn to produce the number, and which model is up to you. Set it in
/plugin configure sawbuck-reins@sawbuck.
native ยท default
A tiny Claude (Haiku) call on your own plan. Zero setup, no key, a fraction of a cent per turn. Falls back on its own if your org blocks model calls.
engine
Hand scoring to a model you choose via SAWBUCK_JUDGE: a free local ollama:qwen2.5:7b (nothing leaves your box), or your own key. Keeps it off your Claude plan.
off
No live steer rate. The check, the debt, and the memory keep working exactly the same.
The judge is a separate, cheap classify call (never your session's main model) that runs in the background, so it never blocks a turn. The deterministic check never calls a model, however the steer rate is scored.
The cue inventory. Past the meter, each turn the judge files your steer into a cue inventory, so the agent picks up your standing preferences as you work. These are leads you review, never rules it adopts for you.