TREFT
The decision card

A sample decision, end to end.

This is the card Treft hands you when an agent’s change is ready: what it claims, what the independent re-run proved, how much of your attention it earned, and the human decision — recorded. Annotated, so you can read it the way a Treft user does.
Illustrative sample · fictional repository · real anatomy

The identities below are fictional — acme/payments-api, the agent atlas-agent, the reviewer sam-hale. Everything else is the shipped product’s own anatomy, vocabulary, and grades, drawn from the same captures used across this site.

Merge decision · github pr 44m ago

Add exponential backoff to webhook retry delivery

Webhook retries now back off exponentially with jitter; 4 files; suite re-run green in the clean room.

READ ×3 large change surface · 3 files real surface
The review tier The card’s first word: how much of your attention this change has earned. GLANCE means verified with a small real surface; READ ×3 means three hunks deserve your eyes and the mechanical rest is named; FULL REVIEW means slow down — and the card says exactly why.
PR by atlas-agent Agent 4 files, +89
Provenance Who authored the change, stated — an agent, a colleague, or Treft’s own engine. The author never grades its own work, so the grade below comes from Treft’s re-run, not from the author’s claim.
TESTS RE-RUN CLEAN-ROOM · L1

The test suite passed on the exact patch that will merge.

The evidence grade In plain words: Re-run clean. Treft built a fresh checkout on a separate machine, applied the exact patch, cut the network, and ran the repo’s own tests. The L1_clean_room code names how far verification went — the grades are published, with the verifier’s measured false-pass rate.
Evidence

vitest — payments-api

27
Pass
0
Fail
1
Skip
2.8s
Duration
Receipts
Tests pass — clean room evidence: L1_clean_room †CLEAN ROOM
No test files changed 0 test files of 4 changed †PATCH
Changes confined to src/webhooks †PATCH
Authored by an agent atlas-agent · via the GitHub App †PROVENANCE
The receipt law Every summary line cites the oracle that earned it — the clean-room re-run, the patch itself, or the provenance record. No cited signal, no summary. A claim Treft cannot back with a receipt is a claim you never see.
↑ HOLD TO SHIP
REQUEST CHANGES REJECT
APPROVED — SHIPPED by sam-hale · squash-merged into acme/payments-api · recorded in the append-only audit trail with the grade and its evidence
The human decision The final act is a person’s. Approve holds to ship; request changes sends the feedback back (on a pull request, as a real GitHub review); reject records why. Nothing ships without your approval — and the approval itself becomes part of the evidence trail.

The three review tiers

Not every diff deserves the same attention. Treft stamps every verified change with a review tier — an honest reading of how much scrutiny the change has earned:

GLANCE

Verified, adequate tests, clean checks, a small real surface.

One look at the one thing worth your eyes.

READ ×3

A few hunks worth attention; the mechanical rest is named.

Read the marked hunks. Skip the noise.

FULL REVIEW

Escalated paths, weak tests, or a wide blast radius.

Slow down. The card says exactly why it earned the scrutiny.

Try it on your own pull request

Connect one repository and Treft grades its open pull requests the same way — the re-run is real, the receipts are yours, and no branch-rule change is required to start.

Verify your first PR — free How verification works

No card · choose one repo · disconnect anytime