localharness
scaffolding that makes a small local model do senior-dev work
prod bugs · one shot
29/47
with retries (best-of-3)
33/47
same harness, best-of-3 — the lean 35B MoE leads
the scalpel toolkit
grep search
find code, don't guess
surgical edit
change lines, not files
full + ranged reads
see the whole file
smart context engineering
real verified window
use actual, not assumed
dynamic in/out split
max output every turn
compress when full
only near the limit
bail if too small
clear warning, no waste
keeps a weak model on rails
forced commit
stop exploring, edit
circling fast-fail
kill loops in seconds
edit over rewrite
no whole-file blowups
rich test feedback
sees why it failed
capability boosters
planning step
map multi-file fixes
best-of-N retry
diverse second attempts
AGENTS.md + skills
repo map + playbook
benchmarked on a live production Laravel app · graded by its own test suite (fail→pass) · a frontier model as the ceiling control