Route learning
How AutoMode learns per project: a kind of task that needed rescue starts on a stronger model next time. It only raises a route and stores no prompt text.
AutoMode keeps a small memory per project. A kind of task that needed rescue in this project starts on a stronger model next time.
How it decides
Each task is filed
After every turn, and after a correction, the task goes into a bucket: its tier and its need.
The outcome is rated
A task is bad when it moved up to Opus after an escalation, or when you corrected or retried it. It is ok when a check passed.
A bucket moves up
Once a bucket has 2 bad tasks and a 40% bad share in 30 days, the cheaper models are left out for that bucket.
What it never does
- It never lowers a route. Learning only raises one.
- It never overrides a pin. Your
/modelkeeps its model. - It never stores prompt text. The file holds task ids, tier and model labels, outcomes and times.
Where it lives
~/.claude/plugins/data/automode/learned-<project hash>.json. It keeps the last 200 tasks. Delete the file to forget a project.
Turn it off
Set "learning": false in config.json, or run /plugin configure automode@automode and set Route learning to off.
node bench/learn-replay.mts replays your own decision records through the learning rules. It calls no model and costs nothing.
Last updated on
Model switches and the prompt cache
Does switching Claude models mid-session waste the prompt cache? It can. The check AutoMode runs before every model switch, and when it holds instead.
Delegated reads
AutoMode's optional read gate: a full read of a large file stops once with a free outline and a cheap Haiku extract tool. How it works and stays safe.