Question 1

Why use a decision model instead of asking my coding agent to self-check?

Accepted Answer

The README's arithmetic is that one check runs about 2,500 tokens, which at ordinary model prices is a cent or more and a few seconds per edit, and the reply is prose you then have to parse and cannot fully trust. A decision model returns a typed answer with a calibrated probability and nothing to hallucinate, at roughly a tenth of a cent and about 300 ms, which is what makes checking every edit on every turn affordable.

Question 2

Will it fight my linter or start enforcing rules I never wrote?

Accepted Answer

No. There are no built-in rules, and nothing is enforced until your instruction files compile into a rubric on the first turn, with the rules quoted back so you can see what was found. Rules a linter could check are deliberately handed to your linter instead. Every verdict names a rule in that rubric and every rule cites the line of your instruction file it came from, so a wrong verdict is something you rewrite rather than debug.

Question 3

One of my rules never fires. Is the tool broken or is the rule bad?

Accepted Answer

Usually the rule. The README says a badly worded rule scores about 0.4 on everything and never fires, which is indistinguishable from silence in the output. The calibrate command scores every rule against twenty real hunks from your own git history and switches off the ones that never fire, while tune has an agent rewrite them. Rules marked for the end of a turn also have no answer after the first of several edits, by design.

Question 4

What data leaves the machine, and do I need Claude Code installed?

Accepted Answer

Changed lines go to TypeSafe under your own key and nowhere else, with no server of the project's in the path, and the README notes retention is between you and TypeSafe because their API takes no per-request retention flag. Separately, rule compilation and tuning use Claude Code by default or Pi when Claude is unavailable, and if neither is present the tool prints the prompt for you to paste into a session instead.

Question 5

At which moments can a rule run?

Accepted Answer

Each rule runs at one of two moments: edit, after each edit, or turn, once at the end against the whole diff. A question like whether an edit added more than was asked has no answer after the first of several edits.

Question 6

How do I exclude files from Abide's checks?

Accepted Answer

Put a .abideignore at the repository root to skip files for every rule, and each rule can also carry a scope of globs, so an API route and a stylesheet get different questions.

Question 7

What does abide audit do to existing code?

Accepted Answer

It judges every file as if it had just been written and reports a table by rule and a list by file. On 33 API routes of a real Next.js app the README reports 12 seconds for about a cent.

Abide

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community