ToranaField notes
Notes from building Torana.
Experiments, architectural detours, and the occasional “how did this become a plugin runtime?”
Can decision models like Jev and Laya route a coding agent?
Laya as a model router across four coding agents: the plumbing worked in ~112 ms, the classifier never cleared its confidence bar, and the cheap model beat the expensive one 10/10 to 8/10 on a small test.
Scan tool output for PII with a local model.
Set up the PII plugin to check for exposed keys, passwords and private data. Compare the local scanners tested, including their misses and false positives.
This was supposed to be a compactor.
How an experiment in compacting tool outputs became a local-first, programmable proxy—with a little architectural scope creep along the way.
Does compacting tool output lower a coding agent's API bill?
A paired DeepSeek V4 Pro experiment: 75 sessions, 4.81 MB of repeated output removed, no measurable saving. Includes the experiment setup, methodology and why prompt caching made the difference.