Prove · Loop Engineering
Loop Engineering
Close the loop on your coding agents so the loops you run improve instead of just running. Loop Engineering is the positioning lens over the machinery that measures a loop, learns from it, and feeds the next run.
What it does for you
Should this be a loop?
Most tasks are not worth a loop yet. Before you build one, the Loop Readiness Wizard on this tab asks the four conditions from the loop-engineering playbook (does it repeat weekly, can a test or type-check or lint or build fail the work, can the agent run what it writes, can the budget absorb the retry waste) plus the two boxes telemetry cannot see: is there a hard stop, and does a human approve before anything irreversible. It returns a good-first-loop, keep-it-manual, or not-enough-to-say verdict with plain reasons. Judgment-call work (architecture, auth or payments, production deploys, vague product work) stays manual no matter how you answer. When the verdict is a good first loop, one click scaffolds it, emit-only.
Where to find it
- Desktop:
http://localhost:4000/, then Prove → Loop Engineering in the sidebar. - Hosted:
repoops.ai/team/loop-engineering - Keyboard: ⌘ K, then type “Loop Engineering”.
The three layers
Harness, loop and graph are three layers of one system, not three competing ideas. RepoOps grades each on its own published rubric, and the three layers panel on this tab reads across them.
- Harness - least privilege, prompt-cache efficiency, the share of your failures the harness owns, and context hygiene.
- Loop - the six loop disciplines this page has always graded.
- Graph - whether subagents stayed inside the scope declared for them, and how much of the spawn tree that check reaches at all.
The three are equally weighted. There is no evidence one matters more than another, and inventing a weighting would be exactly the confident arithmetic these scores otherwise refuse.
An unmeasured layer is shown as unmeasured. It renders hatched and labelled, carrying the reason its own score gave, and it is never counted as zero. Below two measured layers there is no overall number at all, because one layer is not a verdict on a system of three. Expect the graph layer to read unmeasured for now: an attested spawn currently records only when a subagent finished, so nothing about concurrency can be computed until a start time is captured.
Endpoints: GET /api/layer-scores for the roll-up, or /api/harness-score and /api/graph-score for one layer. Every response ships the rubric it was scored against.
Built vs. planned
Shipped and live in the Prove category on both the desktop app and the hosted dashboard. The Loop Readiness Wizard and the Loop Readiness Advisor run on the desktop app only; the hosted tab at /team/loop-engineering is the six-stage rollup plus loop-contract coverage.
Last updated