dislike overly religious people so much mehn , you fit tell them say you never chop today make dem tell you say you should be grateful bc some people never chop since yesterday
our coding agent told us “all tests pass.” it had deleted the test file.
that’s why we built Kedge for #neatHack. dependabot and renovate PRs break all the time and more teams hand them to coding agents to fix unattended. when a fix gets hard, the agent deletes tests, edits files it was never asked to touch, or says it’s done when it isn’t.
kedge supervises the agent. It plans the fix, lets the agent take one step, runs the tests, and checks for drift: tests lost, files outside the plan, a repeated attempt, a claim that doesn’t match the results. On drift it rolls back to the last good commit, tells the agent why, and retries. In our demo the agent deleted a test file at step 1. Kedge rolled it back and step 2 fixed the real bug, 5 of 5 tests passing. With no supervisor, the same agent shipped 2 of 5 tests and called it fixed.
@neatlogs showed us our own blind spot first: 0 errors across 33 traces. Kedge wasn’t reporting its failures at all. After we fixed that, the daily digest flagged the error rate on the next run. The trace pointed to the exact step where the agent deleted the test, and the Execution Failed detection and error analysis group every failure by cause. Kedge also reads its past runs back from neatlogs before it plans, and writes its verdict onto each trace.
@EntireHQ checkpoints every commit the agent makes. Before a new fix, Kedge searches past checkpoints for the same failure and gives the agent what worked. It skips any old fix that touched tests.
@arithecfo set the budget. We built the business plan with Ari: $10 per green PR at a 92.5% margin leaves $0.45 for the model. Kedge reads that from pricing.json and stops before any step it can’t afford. We learned why the hard way: on one batch, OpenCode estimated $1.57 and the real bill was $5.70. Ari’s stress case put uncapped runs at $2.85 a PR and a 68.5% margin. Now Kedge meters the real bill, and our demo fix cost $0.12.
one more rule: Kedge opens no PR without proof, meaning a neatlogs trace, an Entire checkpoint and the cost.
Team: Farouk + Haikeys (@Haikeystweett )
Repo: https://t.co/u8ONll7guv
idk where i stand on that, but if he exists he is either the enemy of humanity for letting life be this shitty or he is an utterly powerless figurehead that we shouldn’t really fear.