Why it mattersRules a linter cannot check ("never let a raw error reach a user", "no premature abstractions") are now enforced in the same turn the agent breaks them, so the human review step catches design mistakes instead of style ones.
The developer coldteadotai published abide, an MIT-licensed hook layer that sits between a coding agent and its file edits and stops the agent from breaking rules it read in AGENTS.md or CLAUDE.md. The repository was created on 2026-09-18 and reached 125 stars in the first thirty hours. It works with Claude Code, Codex and OpenCode.
What it reads and when it fires
On session start, abide hashes the instruction files in the repo (AGENTS.md, CLAUDE.md and any others the agent loads) and, if they changed, asks the agent to compile them into .abide/rubric.json. The rubric holds one entry per rule, with a scope of file globs and a phase: edit for rules checked after each individual edit, turn for rules that can only be judged against the whole diff at the end of a turn. Rules a linter could check are handed to the linter instead. After each edit, abide sends the changed lines and one rule at a time to Jev, TypeSafe's decision model, and gets a calibrated probability back. Above 0.8, the agent is told which rule broke and repairs the file in the same turn. Between 0.5 and 0.8, the score is logged but the agent keeps going. Below 0.5, nothing happens.
The numbers the project publishes
The README states that in 93 real Claude Code sessions replayed against two repositories' own AGENTS.md files, "1 in 13 turns" broke a rule no static linter could catch. The replay covered 1,256 edits and 147 turns and cost 22 cents in total. Jev flagged 39 edits and 15 turns; the project says an independent reviewer confirmed 10 of the edit flags and 11 of the turn flags. A check on the abide repository's own 13 rules is 1,000 to 1,600 input tokens, which the project prices at $0.00004 to $0.00007, and takes 300 ms for Jev's answer, or one second including the hook script and Node startup. A turn of 15 edits costs a tenth of a cent. These figures are the tool's own, measured on 2026-09-18 against the direct TypeSafe endpoint.
The design that makes this affordable
An ordinary LLM check of every edit against every rule was never economic. The project says a check is 2,500 tokens, which at typical model prices is a cent or more and a few seconds per edit, and free-text output the caller had to parse and could not fully trust. Jev is a decision model. It answers a single typed question with a calibrated probability, so there is nothing for the model to invent and no prose to parse. The project cites this as the change that makes edit-by-edit checking practical: the same check, in 300 ms, at a fraction of a cent, for 200 edits a day.
Two hundred edits a day is where the shape of code review changes. When the design rules a repository actually cares about are enforced in the same turn they are broken, the reviewer stops correcting style and starts reading intent. That shifts what a senior engineer's time is spent on, and closes the gap between a change that reads fine and a change that is actually right. The catch is only as good as the rubric: abide's calibrate command scores every rule against twenty real hunks from the repository's git history and switches off any rule that never fires, and tune has the agent rewrite those. A rule you cannot phrase clearly enough for Jev to answer is a rule your teammates also cannot follow.
Source
Primary source: coldteadotai/abide on GitHub.
Source: coldteadotai on GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
