| Route | What the repo owner does |
|---|---|
| green-is-enough | merge it on a green tick, nobody needs to read it |
| tests-are-enough | merge it once a test covers the change |
| ai-review-is-enough | run an AI review; if it finds nothing, merge it |
| human-required | put it in front of a person before it merges |
| human-plus-author | a person reads it with the author walking them through |
Jev answers eleven questions in one call. Code does every piece of arithmetic, and the output is a route rather than a merge.
- -Title, description, the file list, and as much of the diff as the budget holds. A patch - too large to send whole arrives as whole files first, then paths and line counts.
-One call to the checks API per pull request, before any Jev call. Three states are - terminal: no questions asked, no route computed, nothing spent. On eighteen - apache/airflow pull requests this skipped ten of them.
-| State | Means | Waiting on |
|---|---|---|
| ci-failing | a check concluded failure, timed out, or wants action | the checks, or the branch |
| ci-pending | a check is still running | nobody, come back later |
| draft | the author marked it draft | the author, who is not asking |
A name that fails on several unrelated pull requests is the check being - broken rather than each change breaking it: a boto3 bump cannot break Postgres - serialization. Five of eight failures were shared that way, and the output says so, - because the state is a fact and never a verdict on the author.
-Every question sees the same pull request. None sees another's answer, so the eleven - cost the tokens of eleven and the wall clock of about one.
-Nine scored, and weighted
-blast radius how far it reaches · security surface auth, secrets, - tokens, input validation · breaking change a caller stops working · - infra surface deploy, images, CI · design decisions judgment calls to - agree with · mechanical one edit repeated · scope creep the diff - exceeds the description · has tests the change is covered · - description quality what, why, and how they checked
-Two labels, never scored
-What kind of change: docs, tests, dependency, config, bugfix, feature or
- refactor.
- What to check first: correctness, security, compatibility, tests, deployment
- or readability.
- Both print beside the route.
Effort and consequence are different questions, and they take different arithmetic.
-Effort: the weighted mean of nine
-All nine, at the weights in questions.yml: blast radius - 0.15, security surface 0.15, design decisions 0.14, mechanical 0.12, breaking - change 0.11, scope creep 0.09, infra surface 0.08, has tests 0.08, description - quality 0.08.
-Averaging is right here, because a diff that is long in one - way and short in three really is a medium read. Mechanical, has tests and - description quality count inverted, so a repeated edit, a covered change and - a thorough description each pull it down.
-Consequence: the max of four
-Four of the same nine, read differently: security surface (auth, secrets, - tokens), breaking change (an existing caller stops working), blast - radius (reaches shared code paths), infra surface (deploy, images, CI).
-The max, never the mean, because a change that is safe - in three ways and dangerous in one is a dangerous change. A score reads the - probability on its top level only: blast radius level 1 is "one module", - which is ordinary work.
-Consequence sets the floor. Effort can raise it and never lower it. Nothing here - says "merge", because a model reported at 67.8% accuracy can decide how much - evidence to demand and cannot be the last gate before main.
-| Route | Enough to merge on | Reached when |
|---|---|---|
| green-is-enough | CI passing | consequence under 0.15 |
| tests-are-enough | CI, and a test that exercises it | consequence 0.15, or effort 0.42 |
| ai-review-is-enough | the above, and an AI review finding nothing | consequence 0.35, or effort 0.62 |
| human-required | a person reads it | consequence 0.60 |
| human-plus-author | a person reads it, author walking through | consequence 0.60 and effort 0.62 |
Jev reads a truncated diff, so a 54-file change can read as one repeated edit. - Size is arithmetic, Python and Go own it, and it only ever raises.
-Grouped by route, with what set the consequence, the questions carrying the effort, - and the single signal that would drop it a route: "a test that exercises the - unconnected-identity path". A tier tells an author nothing to do. That line does.
-