adhere

^{A linter for non-deterministic rules. Powered by Typesafe}

A linter for rules a normal linter cannot check, such as "Business logic should live in services", or "modules should expose lots of functionality behind a small API". Define rules as formatted markdown:

---
description: A third party API going offline should not also take our app down.
---

Our APIs should never be coupled to some third party API on a critical path. The user facing
side should be unaffected if it goes offline, and ingestion should happen in a separate asynchronous process.

## Should

```ts
import googleMaps from "./lib/maps"
import db from "./lib/db"
import redis from "redis"
import cron from "node-cron"

// asume this is called by /api/restaurants/:zipcode
export async function findRestaurants(zipCode: number) {
  const results = db.restaurants.find({ where: { zipCode }})
  return results
}

export async startWorker() {
  cron("0 0 * * *", async () => {
    const since = Number(await redis.get("maps:sync:timestamp"))
    const items = await googlemaps.findAll({ since, limit: 100 })
    const syncItem = item => db.restaurants.upsert(item)
    const next = items.sort((a, b) => b.id < a.id).at(-1)?.timestamp
    const touch = () => redis.set("maps:sync:timestamp", next ?? Date.now())

    return Promise.all(items.map(item => syncItem(item))).then(touch)
  })
}
```

## Never

```ts
import googleMaps from "./lib/maps";
import db from "./lib/db";

// asume this is called by /api/restaurants/:zipcode
export async function findRestaurants(zipCode: number) {
  // what happens when google maps goes down? how about latency? how about costs?
  const results = await googleMaps.findAll({ where: { zipCode }, limit: 100 });
  return results;
}
```

Your rules are parsed, then adhere asks Jev (TypeSafe AI's System One model) whether the file breaks each rule, gets a calibrated probability per rule, and reports the ones above a threshold with the section of the file Jev points at. The report uses the same frame as vp lint.

Here's a demo of what the output looks like.

Demo 🎬

Presets for TypeScript, React, security, Effect, and alchemy are included, and run without setup:

TYPESAFE_API_KEY=xxx npx @drkmttr/adhere lint --preset typescript

Rules a normal linter can check exactly, such as a banned import or a type error, belong in that linter. While it lints, adhere asks Jev whether a normal linter could check each of your rules. adhere validate lists the ones it thinks belong in a regular linter, and detects contradictions between your rules.

Install

curl -fsSL --create-dirs -o ~/.local/bin/adhere \
  https://github.com/darkmatter/adhere/releases/latest/download/adhere-$(uname -s)-$(uname -m)
chmod +x ~/.local/bin/adhere

adhere is a single executable for macOS and Linux (arm64 and x64) and for Windows (x64), and needs nothing else installed, Node included. Any directory on your PATH works in place of ~/.local/bin. On Windows, download adhere-Windows-x86_64.exe from the latest release.
Run the command again to update. download/v0.14.0 in place of latest/download fetches that version.
To pin the version in a JavaScript repo, for CI or scripts, add @drkmttr/adhere as a dev dependency and run npx adhere. That adhere starts through Node, or through Bun under bunx.
adhere sends each file it judges to Jev at api.typesafe.ai, authenticated with a TypeSafe AI API key.

Login

adhere login             # prompts for the key, masking what you type
adhere login < key.txt   # reads it from stdin instead
adhere logout            # deletes the saved key

The key is saved to ~/.config/adhere/credentials.json, or under $XDG_CONFIG_HOME when that is set, readable only by you.
TYPESAFE_API_KEY, when set, takes precedence over the saved key, so CI can pass a key without a login.
The key is read only when a request is about to be sent. A run where every file is cached needs no key and no network.

Quick start

adhere init        # .adhere/config.ts, two example rules, and the dependency
adhere login       # save your TypeSafe AI API key, once
adhere validate    # check your rules' wording, and ask Jev whether any contradict
adhere lint        # audit the working directory
git add .adhere    # commit the rules, and the judgments they cost

Init

adhere init creates a config file, some example rules, and adds adhere as a dev dependency.

It is safe to rerun: existing files are reported as skipped, and only --force overwrites them.

Generate rules from your repo

Let an agent write your first rules with you. The adhere-setup skill walks it through setting adhere up in your repo:

It reads where your conventions live: AGENTS.md, CONTRIBUTING.md, docs and ADRs, your lint config, recent review comments, and the code others copy from.
It drafts the conventions a normal linter cannot check as rules, each with real code from your repo, and sets aside the ones a linter could check.
It adds the presets that fit your stack, and rules from your organization's shared repo, if you have one.
It shows every candidate in one list, with its evidence, and imports the ones you choose.
It validates them, walks you through the first lint and what it costs, and adds adhere to CI.

It stops to ask you at each of those decisions, and never asks for your API key.

skills add darkmatter/adhere   # install adhere's skills for your agent

Then ask your agent to set adhere up.

Install rules from other repos

An organization's rules can live in one repo's .adhere/rules/, and other repos copy them in with adhere install:

adhere list darkmatter/standards                        # each rule there, with its description
adhere install darkmatter/standards                     # every rule
adhere install darkmatter/standards/data                # the data topic
adhere install darkmatter/standards/data/brand-ports    # one rule
adhere install darkmatter/standards#v3                  # every rule, as tagged v3

Copies land in this repo's .adhere/rules/org/repo/. So data/brand-ports from darkmatter/standards is .adhere/rules/darkmatter/standards/data/brand-ports/, the rule darkmatter/standards/data/brand-ports, and two repos' rules never collide.
The copies are the repo's own rules from then on: commit them, edit them, or start new rules from them. Nothing tracks where they came from.
A later install skips a rule already there, as init does. --force replaces its directory, edits and all.
A rule's whole directory is copied. The source's config, its inline rules, and its cache are not.

Read what you install. A RULE.ts is copied with the helpers beside it, and runs whenever lint does, in CI too. adhere list names a RULE.ts without its description, since reading it would run it.

The source is cloned with git from https://github.com/org/repo.git, so a private repo needs git's credentials for GitHub, as gh auth setup-git sets up. To clone over SSH instead, let git rewrite the URL: git config --global url.git@github.com:.insteadOf https://github.com/.

Creating a shared repo

adhere init --shared org/repo scaffolds a repo whose rules other repos copy in with adhere install, instead of one that lints itself. It writes the two example rules and no config, since install copies only rule files, and with them:

File	What it is for
`README.md`	Says what the repo is for, how to install its rules, and how to add one.
`.github/workflows/adhere.yaml`	Runs `adhere validate` on every push to main and every pull request, with the `TYPESAFE_API_KEY` secret. A pull request from a fork gets no secrets, so validate refuses there.
`alchemy.run.ts`	An alchemy stack that creates the GitHub repo, or adopts it, and sets that secret from `TYPESAFE_API_KEY` when deployed with `npx alchemy deploy`.
`package.json`, `.gitignore`	alchemy and effect, pinned to versions that work together, which init installs, and a `.gitignore` for them and alchemy's state.

The repo is public unless you answer yes when init asks, at a terminal, whether to make it private. Other repos then need git's credentials for it to list and install its rules. Without a terminal it is public: change visibility in the stack to make it private.

Check for contradictions

Two rules contradict when no code can follow both, such as one that says errors must be thrown and one that says they must be returned. Each rule reads fine on its own, so a contradiction shows up only as code that breaks one rule or the other whatever you do. adhere validate finds them before lint does:

adhere validate                  # your rules
adhere validate --preset effect  # your rules beside a preset's

Found 1 contradiction among configured rules:

1. No code can follow both (0.91):
   - errors/throw-tagged-errors (/repo/.adhere/rules/errors/throw-tagged-errors/RULE.md)
     A function that fails must throw a tagged error, never return an error value.
   - errors/return-results (/repo/packages/api/.adhere/rules/errors/return-results/RULE.md)
     A function that fails must return a Result, never throw.

Resolve by editing one rule, narrowing a nested rule's scope, or using the same rule id when the nested rule is meant to shadow the root rule.

It asks Jev about each pair of rules that apply to the same files, so it needs the API key, as lint does. A contradiction fails it, with exit code 1, so it can gate a pull request that adds a rule, as the workflow adhere init --shared writes does.

It also runs two checks that are advice and never fail it:

Wording: each rule worded otherwise than the rule writing tips recommend, with its file and how.
The linter check: while lint judges a file against one of your rules, it also asks Jev whether a linter or type checker could decide the rule exactly, on 10 files per rule. A rule flagged on 7 or more of them probably belongs in a regular linter. One flagged on 3 to 6 reads differently from file to file, so its description should say more precisely what it applies to. Preset rules are left out.

Two rules with the same id are never compared: a nested rule that shares an id shadows the other on purpose.

Live evals

Try a change to a rule on your own code before you rely on it: run the current rule and the changed one side by side, and compare what each flags.

Copy the rule's directory under a new id:

cp -r .adhere/rules/data/brand-ports .adhere/rules/data/brand-ports-next

Give the copy level: warning in its front matter. Its findings show in amber and never fail the run, so the experiment can sit in the repo, and in CI, while you watch it.
Change the copy: its description, its examples, its appliesTo or excludeIf, or what it reads. To try giving Jev more to read, make the copy a RULE.ts with an appendState hook.
Run adhere lint, or adhere lint --filter '<glob>' to try it on part of the repo first, and compare the two rules' findings on the same files.

Judge the two by recall, the share of real violations each catches, and precision, the share of its flags that are real. Check the findings by hand, or with the adhere-fix skill, which verifies each one against its rule and the code. Then keep the better version under the original id, and delete the other.

The copy costs nothing until you change it. Judgments are cached by the rule's text, not by its id, its level, or its threshold, so a copy that reads the same as the original answers from the original's cache. Each change after that judges the copy, and only the copy, once per file.
Keeping the winner is free. Moving the winning text to the original id reuses its cached judgments.
Counting each rule's findings takes a pipe: adhere lint --yes | grep -c 'data/brand-ports-next:'.

Writing good rules

We've evaluated how Jev reads a rule, on the presets and on other repos, to catch the most violations with the fewest false positives. The eval has every study; these are the lessons for writing a rule.

Rule writing tips

Say "must" and "never", in the description and in the headings over the code, or "should" and "should not" for a guideline. Rewording the presets' descriptions this way, with a bad example per rule, was the largest gain of any study: at 0.8, as many caught, and 1 wrong where there had been 5 or 6 (study).
Give one example of each kind: one must block and one never block. A bad example helped every wording, and at 0.9 caught 24 violations with none wrong, where there had been 17 (study). Either kind alone did worse, and three of each did no better, for 1.58 times the tokens (study).

adhere validate lists each rule worded otherwise.

What else the studies found

Lesson	What the study found
"Should" is for people, not a weaker check.	Rules written with "should" and with "must" scored nearly the same (study).
A wrong finding is usually the rule's fault.	Above 0.8, 31 of 35 wrong findings repeated within a rule: a scope its words reach past, or the same misreading again. Fixing the wording fixes them together. Below 0.8, one-off misreads grow, to about one finding in six at 0.6 to 0.7 (study).
Narrow a rule in its description.	Jev scores an `excludeIf` near 0.5 for real and false findings alike, so it thins a rule's findings about as much as a higher threshold (study). Narrowing the description moves the judgment itself: scoping alchemy's idempotent-delete rule to providers' delete handlers dropped its findings elsewhere from 0.67–0.88 to 0.08–0.21, with nothing real lost (study).
Beware rules that turn on what the file does not show.	The weakest preset rules hinged on facts outside the file: whether a client has a default timeout, whether a script is an entry point, whether a key is public (study). Warning below a context of 0.7 marked one finding in five, and those were real about half the time, where the rest were real three times in four (study). Make such a rule a warning, or give Jev the fact with `reads`, an `appendState` hook, or an `@adhere` note.
Jev believes comments.	Eight of nine real violations of the idempotent-delete rule carried a comment wrongly saying they were fine, and scored 0.31 to 0.62; without comments, all nine scored 0.81 or more (study). That is why adhere takes comments out. A fact Jev needs, such as an exemption, goes in an `@adhere` note, which stays (study).

Commands

Command	What it does
`adhere lint`	Audit the working directory. See Running lint.
`adhere validate`	Check the rules' wording, and ask Jev whether any contradict. See Check for contradictions.
`adhere init [--force]`	Scaffold `.adhere/config.ts` and two example rules. See Init.
`adhere init --shared org/repo`	Scaffold a repo of rules other repos install. See Creating a shared repo.
`adhere list org/repo`	List the rules in another repo's `.adhere/rules/`. See Install rules from other repos.
`adhere install org/repo`	Copy them into this repo's `.adhere/rules/org/repo/`.
`adhere login`, `adhere logout`	Save or delete a TypeSafe AI API key. See Login.
`adhere skill [docs\|setup\|fix]`	Print an agent skill. See Agent skills.

Bare adhere prints the help. adhere <command> --help lists a command's flags, and adhere --completions <shell> prints a completion script.

Reading the report

A finding looks like this:

  × data/brand-ports: A port must be a branded, range-checked integer, never a bare number.
    confidence 0.93 · context 0.88
   ╭─[src/server.ts:6:3]
 1 │ import { Effect } from "effect";
 2 │ import { listen } from "./listen.ts";
 3 │
 4 │ export const serve = Effect.gen(function* () {
 5 │   const host = process.env.HOST ?? "localhost";
 6 │   const port: number = Number(process.env.PORT ?? 3000);
   ·   ──────────────────────────────────────────────────────
 7 │   yield* listen({ host, port });
 8 │   yield* Effect.log(`Listening on ${host}:${port}`);
 9 │ });
   ╰────
  hint: const Port = Schema.Int.pipe(
          Schema.check(Schema.isBetween({ minimum: 1, maximum: 65535 })),
          Schema.brand("Port"),
        );

  confidence  0–0.74 < 0.8 < 0.85–1  Jev's probability that the file breaks the rule
  context     0–0.49 < 0.6 < 0.70–1  its probability that the file shows enough to decide

Found 1 error.
42 files, 3 judged, 39 cached.

Part	What it tells you
`×` or `⚠`	An error, or a warning, in amber, from a rule whose `level` is `warning`. Only errors fail the run.
`data/brand-ports`	The rule's id, then its description. A preset's rule has the preset's name first, as in `effect/basics/gen-for-sequencing`.
`confidence`	Jev's probability that the file breaks the rule.
`context`	Jev's probability that the file shows enough to decide that.
The excerpt	The section of the file Jev points at. The line it names is underlined, or the lines, when it names several, are marked with a bar beside the code.
`hint:`	The rule's code that must be written. A rule with only code that must never be written shows that code instead, labeled `never:`.
The legend	Under the last finding: each score's threshold, between its range in red and its range in green.

Score colors. On a terminal, each score is colored by where it sits around its threshold:

If you're near the threshold (orange), consider adjustments.
Red is below the threshold, so lint never reports it: only the legend shows that range.

Low context. Jev sees one file at a time. So we also ask whether the file shows enough to decide the rule at all. When context is below 0.6, it means Jev thinks it needs more info, which you can add with an @adhere comment. You may also see a hint:

  warning: this file may not show enough to check this rule
     help: add what the code relies on outside this file as a note Jev reads:
           // @adhere <the fact>, and how you know it

On the eval, most findings with this warning were false, where about one in five of all findings was (study). Check what the code relies on outside the file before acting on it.
If you feel the threshold is too low/high, override it using sufficiencyThreshold in the config.

Rule format

A repo's own rules live in .adhere/rules/, organized just like skills:

.adhere/
  config.ts           optional
  cache/              commit it
  rules/
    data/             a topic: a directory that holds no rule
      brand-ports/
        RULE.md       the rule data/brand-ports
      columns/
        RULE.ts       a rule written in TypeScript
        schema.ts     a helper it imports, not read as a rule

The directory's path is the rule id: .adhere/rules/data/brand-ports/RULE.md is data/brand-ports.
A rule can be markdown RULE.md, or typescript RULE.ts. A directory with both refuses the run.
Files not named RULE are ignored.

Rules as Markdown files

A RULE.md is front matter, then a body that holds the rule's code:

---
description: A port must be a branded, range-checked integer, never a bare number.
threshold: 0.8
---

Why: a bare `number` accepts 70000 and -1.

## Must

```ts
const Port = Schema.Int.pipe(
  Schema.check(Schema.isBetween({ minimum: 1, maximum: 65535 })),
  Schema.brand("Port"),
);
```

## Never

```ts
const port: number = Number(process.env.PORT);
```

Front matter	Meaning
`description`	Required. One sentence saying what code must be, and never be.
`threshold`	The rule's own cutoff, in place of the config's.
`level`	`warning` reports the rule's findings as warnings, which do not fail the run. Use it for a nit, or for a rule that tends to flag code wrongly.
`tests`	Rules skip tests by default. `only` is for a rule about tests, which judges nothing else. `include` is for one that holds in tests as well.
`appliesTo`, `excludeIf`	Where the rule applies. See Scoping a rule.
`reads`	What Jev reads beside the code. See What a rule reads.

Heading in the body	The code under it
`## Must`	must be written
`## Never`	must never be written, which is what a violation looks like
`## Should`, `## Should not`	the same, for a guideline

A rule needs code under at least one heading.
A rule is a requirement ("must", "never") or a guideline ("should", "should not"), and cannot mix the two.
A rule with only a never block suits a rule with no single correct form to show, such as a hand-rolled retry loop or an error caught and dropped.
Neither goes in the other's place: code that must never be written, under must, reads to Jev as the pattern to follow.
Prose around the code is the rule's details: Jev reads it after the description, so use it to say why the rule holds or where it does not apply. The report shows the description alone.

Scoping a rule

appliesTo and excludeIf say where a rule applies, apart from what it asks for. Each is a list of descriptions of code: a JSON array on one line in front matter, or an array in a config or a TypeScript rule.

---
description: Structured variants must be Schema.TaggedClass members of a Schema.Union, never hand-written object types joined by a tag field.
appliesTo: ["a declared union type or schema"]
excludeIf: ["a type that mirrors a third-party format whose tag key that format fixes, such as a Slack Block Kit block"]
---

Key	Jev is asked, of the code that breaks the rule	A finding stands when
`appliesTo`	whether any of it is as described	Jev says yes to every one
`excludeIf`	whether all of it is as described	Jev says yes to none

The questions are about the code that breaks the rule, not the whole file, so a file with a real violation beside code an excludeIf describes keeps its finding.

A yes is a probability above 0.5. At --log-level debug, lint logs each finding its matchers dropped, with their scores.

Narrowing the description often works better. In the preset study, Jev scored an excludeIf near 0.5 for real and false findings alike. See Writing good rules.

What a rule reads

Jev judges a file by its code alone. A rule that turns on something the file does not show can name it in reads, and adhere gives it to Jev beside the code, for that rule only. It is a JSON array on one line in front matter, or an array in a config or a TypeScript rule.

---
description: A function from one of this repository's own packages, which `workspacePackages` names, must be called through the package's namespace, never imported by its own name.
reads: ["workspacePackages"]
---

## Must

```ts
import * as Orders from "orders-core";
```

## Never

```ts
import { parse } from "orders-core";
```

There is one name so far:

Name	What Jev reads
`workspacePackages`	The workspace's packages, each one's name with its directory, as in `{ "orders-core": "packages/orders" }`. A package that is not in it is an installed one. It is for a rule about whose package an import is of: the repository's own, or one installed from a registry.

Each name is the key Jev reads it under, so the description can name it in backticks, as the questions name code.
A name adhere does not have refuses the run, so a misspelled one does not go unnoticed.
What adhere has no name for, such as a file of the repository's, a rule in TypeScript adds with appendState.

More on requests and where the packages come from

The packages are read on every run, when a rule reads them, from the working directory's package.json, whose workspaces npm, Bun, and Yarn read, and from pnpm-workspace.yaml's packages.
In a pattern, * is any directory and ** any depth of them, and one that starts with ! leaves out what it matches. Each matched directory's package.json gives the name, and its directory is given from the working directory.
Nothing need be installed, so a run in CI reads the same packages as one on a laptop.
A working directory that is not a workspace's root has none.

Rules as TypeScript files

A rule's directory can hold it as a RULE.ts instead, which default-exports defineRule({...}), taking the fields a config's inline rule does. What TypeScript adds is appendState, a hook on what Jev reads for the rule:

// .adhere/rules/data/columns/RULE.ts
import { defineRule } from "@drkmttr/adhere";

export default defineRule({
  description:
    "A query must name only columns its table has in `schema`, and never a column it lacks.",
  must: 'db.select("id", "email").from("users")',
  never: 'db.select("mail").from("users")',
  appendState: async (state, file, Bun) => ({
    schema: await Bun.file("db/schema.sql").text(),
  }),
});

The hook runs right before each request about the rule goes out. It gets the request's state, the file the request is about, and Bun's API, and what it returns is spread over the state.
The description can name a key the hook adds, in backticks, as the questions name code.
A RULE.ts is imported, as a config is, so its code runs on every lint.
A rule with a hook goes to Jev in requests of its own, so only that rule reads what its hook adds. Each of those requests sends the file again.
What the hook reads outside the file is not part of the cache: when db/schema.sql changes, a file already judged is not judged again until it or the rule changes.

Where rules live

Rules in the root .adhere/rules/ apply project-wide.
A nested .adhere/ scopes its rules to the directory that contains it. A rule in packages/api/.adhere/rules/data/brand-ports/RULE.md has the same id, data/brand-ports, but applies only to files under packages/api/.
If a nested rule has the same id as a root rule, the nearest containing .adhere/ shadows the less-specific rule for that subtree. Outside that subtree, the root rule still applies.
Nested .adhere/ directories are not discovered under node_modules/, dist/, and the other skipped directories.

A repo that would rather keep its rules with the rest of its documentation can point rules at a directory such as docs/adhere/:

export default defineConfig({ presets: ["effect"], rules: "./docs/adhere" });

That directory holds its rules as .adhere/rules/ does, a directory per rule, and they apply project-wide.

Config

A config file is optional. Without one, adhere reads the rules in .adhere/rules/ and any --preset. A config names presets, sets the model and thresholds, or gives rules inline. It sits in the working directory of the repo being audited, at one of these paths (keep one):

.adhere/config.ts, the default, next to the rules
adhere.config.ts
.adhere.config.ts

import { defineConfig } from "@drkmttr/adhere";

export default defineConfig({
  model: "jev-latest",
  threshold: 0.8,
  sufficiencyThreshold: 0.6,
  presets: ["effect"],
  exclude: ["**/generated/**"],
  overrides: { "effect/basics/instrument-with-pipe": "off" },
  rules: {
    "data/brand-meaningful-primitives": {
      description:
        "A primitive with semantic meaning, such as an id, email, URL, port, or count, must be a branded schema.",
      must: `
const UserId = Schema.String.pipe(Schema.brand("UserId"))
type UserId = typeof UserId.Type
`,
      never: "type UserId = string",
      threshold: 0.8,
    },
  },
});

Key	Default	Meaning
`model`	`"jev-latest"`	The model id sent to TypeSafe.
`threshold`	`0.8`	Report a rule when Jev's probability is above this.
`sufficiencyThreshold`	`0.6`	Below it, a finding warns that the file may not show enough.
`presets`	none	Built-in rule sets, or their topics. See Presets.
`rules`	`.adhere/rules/`	Rules inline, as above, or a directory (see Where rules live). Either replaces `.adhere/rules/`: none of its rules, root or nested, is read then.
`exclude`	none	Globs, relative to the working directory, of files no rule judges.
`overrides`	none	Settings for a rule by its id. See Overrides.
`includeComments`	`false`	Send every comment to Jev. See Comments.

An invalid shape refuses the run.
A config can import other files by relative path, but no packages besides @drkmttr/adhere: the executable does not resolve node_modules.

Presets

Preset	Rules	For	What it asks for
`typescript`	11	any TypeScript project	Data from outside checked at runtime, invalid states unrepresentable, errors never swallowed, resources released on every path, arguments not mutated, and tests that assert outcomes and stand alone.
`react`	10	components and hooks	Logic for an event in its handler rather than an effect, effects that clean up and ignore stale fetches, `useSyncExternalStore` for outside stores, state that neither copies props nor contradicts itself, and Server Functions and Server Components that guard what crosses to the client.
`security`	4	any project	Secrets kept out of logs, error messages, and responses; parameterized SQL; no untrusted input in shell commands, `eval`, or file paths; and secrets compared in constant time. From OWASP's cheat sheets.
`effect`	16	code written with Effect	Steps sequenced with `Effect.gen`, instrumentation attached with `.pipe`, config read through a service and validated, secrets redacted, branded primitives and tagged unions, defects kept apart from typed errors, services built by layers, and tests with their own layers and `TestClock`.
`alchemy`	41	code that deploys with alchemy	Where Config and bindings are read, which resources keep their data, how secrets stay out of bundles and logs, authorization on public URLs, migrations, durable workflows, and custom providers.

Name presets in the config's presets, or on the command line with --preset, repeated or separated by commas, as in --preset effect,alchemy, in which case the config file is optional. Presets named in both places apply together.
A topic, one of a preset's subdirectories, is a preset of its own: --preset effect/basics applies only the rules under presets/effect/basics/, and alchemy/secrets only alchemy's secrets rules. security has no topics.
To change a preset's rule without copying it, use Overrides.

What every preset does

It leaves out conventions a linter checks exactly. The notes on each preset say which linter checks them.
A preset rule says "must" only where its source makes a requirement, and "should" where the source gives advice.
A topic's rules keep the ids they have in the whole preset, so a topic and its preset share cached judgments, and naming both applies each rule once.
Where effect or alchemy had a rule that typescript or security has, that one is kept and the other is gone. typescript's async/network-calls-have-timeouts replaced effect's basics/external-calls-are-resilient, and security/parameterized-queries and security/no-secrets-in-output replaced alchemy's data/parameterized-sql and secrets/never-logged-returned-or-output. A project on Effect or alchemy names typescript and security too, for those rules.

Notes on typescript

Source. The TypeScript Handbook and the Google TypeScript Style Guide.
Left to a linter. What typescript-eslint and ESLint check exactly: unhandled promises (no-floating-promises), throwing non-errors (only-throw-error), any (no-explicit-any), exhaustive switches (switch-exhaustiveness-check), a lost cause (preserve-caught-error), and empty catch blocks (no-empty).
Warnings. Four of its rules report warnings. Three are advice that code often has reason to set aside: assertions that claim only what the code established, types derived rather than restated, and independent awaits run concurrently. The fourth, that HTTP requests carry a timeout, was right on fewer than half its findings above 0.8 on the eval: many of the rest were calls through a client with a default timeout that the file does not show.
Tests. testing/assert-outcomes and testing/independent-tests say tests: only: they judge tests and nothing else.
Gone. Two rules the eval found mostly wrong: that tests use fake time rather than real waits, and that importing a module does no work.

Notes on react

Source. react.dev, mostly You Might Not Need an Effect and Choosing the State Structure.
Left to a linter. What the React Compiler's lint rules check, in eslint-plugin-react-hooks's recommended set and in Oxlint: pure render (purity), props and state never mutated (immutability), refs not read during render (refs), no setState in render or synchronously in an effect (set-state-in-render, set-state-in-effect), which covers state derived in an effect, and no component defined inside another (static-components).
Files. Its rules judge .ts files as well as .tsx, and apply only to components, hooks, effects, and Server Functions, so other files cost their checks and find nothing.
Warnings. server/no-private-data-to-client reports warnings: a file does not show whether the component it passes a record to is a Client Component. So does state/store-ids-not-copies, which also flags state that copies a string or an option from a fixed list, where a copy cannot go stale.

Notes on effect

Source. The effect-solutions docs and the effect/platform docs.
Tests. Three rules say tests: only, so they judge tests and nothing else: testing/test-clock-for-time, services/fresh-layer-per-test, and config/tests-provide-values-directly.
Guidelines. One rule says "should": a test of code that consumes config provides it through a layer.
Domain types. The data rules judge the types services exchange, return, or store. A shape that mirrors a payload until it is mapped, and a view's props or state, are out of their scope.
What the rules allow. Config validation accepts Config.mapEffect as well as Config.schema, and the variants rule does not rule out a switch.
Gone. The rule that a command handler only parses input: the docs show that pattern but do not ask for it. The rule that test layers are in memory, which flagged tests of real services: none of its findings labeled on the eval or on darkmatter/agents was real. And the rule that service operations have no requirements, which leakingRequirements below checks exactly.

Left to a linter. Effect's language service, @effect/tsgo on TypeScript 7, checks seven conventions exactly, which the preset checked through 0.7. Most are off by default. npx @effect/tsgo setup installs it, and these lines in its plugin options in tsconfig.json turn them on:

"diagnosticSeverity": {
  "strictEffectProvide": "warning", // layers are provided once, at the entry point
  "leakingRequirements": "warning", // service methods have no requirements
  "effectFnOpportunity": "warning", // a named function returning an Effect uses Effect.fn
  "preferSchemaOverJson": "warning", // JSON is decoded with Schema, not JSON.parse
  "nodeBuiltinImport": "warning", // platform services, not node: builtins
  "globalFetch": "warning",
  "globalFetchInEffect": "warning",
  "processEnv": "warning",
  "processEnvInEffect": "warning",
  "globalRandom": "warning", // the Random service, not Math.random
  "globalRandomInEffect": "warning",
  "globalErrorInEffectFailure": "warning", // tagged errors, not Error
  "extendsNativeError": "warning"
}

npx @effect/tsgo diagnostics --project tsconfig.json runs them in CI. Bun globals, which the old platform rule also ruled out, need a lint rule of their own, such as no-restricted-globals.

What the linter misses. leakingRequirements sees a requirement in an operation's type, but not a service built by a factory function that takes its dependencies as arguments, whose types have no requirements to find. The preset's services/dependencies-through-layers asks for that.

Notes on alchemy

Source. alchemy's docs and blog.
Warnings. providers/idempotent-delete reports warnings. Whether a delete fails on a resource that is already gone is a fact about the API, which the file does not show, so the rule also flags deletes of APIs that succeed anyway. See the study.

Overrides

A config's overrides changes a rule by the id a report names it with, a preset's with the preset first, without copying its text into rules:

export default defineConfig({
  presets: ["effect", "alchemy"],
  overrides: {
    "alchemy/providers/idempotent-delete": { level: "warning", threshold: 0.9 },
    "effect/basics/instrument-with-pipe": "off",
  },
});

Value	Effect
`"off"`	drops the rule
`"warning"`, `"error"`	sets its level
`{ level, threshold }`	sets either, or both

An id that no preset or project rule has refuses the run, so a typo does not go unnoticed. A rule of a preset the run does not name is accepted.

Precedence

For the model and the threshold, highest first:

--threshold on the command line
the config file
presets, in order (a later preset wins)
the defaults, jev-latest and 0.8

For one rule:

A rule's own threshold beats all of the above for that rule.
An entry in overrides beats the rule's own threshold and level.
A rule in rules replaces a preset rule with the same id.

--sufficiency-threshold beats the config's sufficiencyThreshold, which beats the default, 0.6.

Running lint

Flags

Flag	What it does
`--preset <names>`	Add built-in rule sets, or their topics. The config becomes optional. See Presets.
`--threshold <0 to 1>`	Replace the config's threshold. Per-rule thresholds still apply.
`--sufficiency-threshold <0 to 1>`	Replace the config's `sufficiencyThreshold`: a finding warns when its `context` is below it.
`--yes`, `-y`	Send the requests without asking first.
`--limit <checks>`	Judge at most that many checks. See Limiting a run.
`--rpm <requests>`	Send at most that many requests a minute.
`--filter <glob>`	Read only the files a glob matches.
`--deny-warnings`	Fail on warnings as well as errors.
`--log-level <level>`	Log the run's work on stderr. See Logging a run.

What gets read

adhere lint reads the TypeScript files under the working directory: .ts, .tsx, .mts, and .cts. It does not read declaration files such as .d.ts, adhere's own config files, or JavaScript.
When the working directory contains agents/, apps/, or packages/, only those trees are read.
Below the working directory, anything under node_modules/, dist/, coverage/, vendor/, e2e/, references/, .adhere/, .agents/, .claude/, .direnv/, .alchemy/, or .vite/ is skipped. The directories above the working directory do not count.
.gitignore is not consulted.
A config's exclude lists globs, relative to the working directory, of files no rule judges, such as generated code. It leaves them out of every run, as --filter '!<glob>' leaves them out of one:
```
export default defineConfig({ exclude: ["**/generated/**"] });
```

Tests are judged only by rules about tests, whose front matter says tests (see Rules as Markdown files). On alchemy, 38% of the findings from rules not about tests were in test helpers and fixtures. A test is:

a .test or .spec file, such as a.test.ts or Button.spec.tsx;
a fixture or test helper named as one, such as fixtures.ts, GitHubHttpFixtures.ts, ledger-copy.fixture.ts, linear-test-helpers.ts, or test-utils.ts;
any file under a test/, tests/, __tests__/, testing/, fixtures/, fixture/, __fixtures__/, or __mocks__/ directory.

Before and during a run

Before it judges anything, lint says on stderr what it found and what judging takes, and asks:

200 files and 14 rules: 2800 checks, 1400 cached.
Judging the other 1400 takes 200 requests to Jev, plus 1 or more for each file with a finding.
Those carry about 1.9 million input tokens: about $0.08 at $0.042 per million, and more for locating findings.
? Send 200 requests to Jev, about $0.08? › (Y/n)

It asks only with a terminal on stdin and stdout, and never when the cache answers every check. --yes sends without asking.
The cost is adhere's estimate of the input tokens times the model's price. Jev charges only for input tokens: $0.042 a million for jev-latest, per TypeSafe AI's models page in September 2026.
A run that stops loses nothing: files finished before it stopped are cached, so running again continues from there.

More on unknown prices and blocked files

For a model adhere has no price for, the plan gives the tokens alone.
The firewall in front of Jev's API can refuse a request whose code reads to it as an attack, with a 403 and an HTML page rather than Jev's JSON. It refuses the same request every run, so that file is skipped, the run goes on, and the report lists each such file with Cloudflare's Ray ID, which TypeSafe AI can look the block up by.
When only the question that finds the line is refused, the file's judgments are kept, and a rerun sends that question alone.

Limiting a run

Three flags bound what a run does:

Flag	Effect
`--limit <checks>`	Judges at most that many checks, taken in path order. The rest wait, and since judgments are cached, the next run with the same limit picks up where this one stopped. `--limit 0` shows the plan and judges nothing.
`--rpm <requests>`	Sends at most that many requests to Jev a minute, evenly spaced, retries included. The plan says about how long they take.
`--filter <glob>`	Reads only the files whose path from the working directory matches, such as `src/` or `/*.service.ts`. Wrap a glob in single quotes so the shell does not expand it. Repeat the flag for more, and start a pattern with `!` to leave out what it matches.

adhere lint --filter 'packages/api/**' --filter '!**/generated/**' --limit 200 --rpm 30

Logging a run

Level	What it logs, on stderr
`--log-level debug`	The config it loaded, how many paths it listed and how many of them it reads, and each request to Jev, with the file it is for, about how many tokens it carries, the HTTP status, how long it took, and Cloudflare's Ray ID. A failed attempt is logged even when a retry hides it.
`--log-level trace`	Also each file as it is read, planned, and answered from the cache, and the first 4000 characters of any error Jev's API answers with.

stdout still carries only the report, so the log can go to a file of its own:

adhere lint --log-level debug 2> adhere.log

Exit codes

Code	When
0	Nothing is reported, or only warnings.
1	An error is reported. With `--deny-warnings`, a warning too.
1	The run refuses, for example on an invalid config or rule file, a missing API key, or an unknown command or flag. A refusal prints its reason.

Comments and suppressions

Comments

Jev reads a file's code, not its comments. adhere takes every comment out before it hashes or sends a file, blanking it so every line keeps its number. The report's excerpts still show them.

Jev takes what comments claim at their word, which hid real violations in the studies: see Writing good rules.

A comment that says @adhere anywhere in it is a note for Jev, and stays. Write one for a fact the code relies on that the file cannot show, once you have checked it, with how you know:

/**
 * Deletes the activity. @adhere DeleteActivity succeeds on a missing
 * activity, so no not-found error needs catching; probed 2026-09-25.
 */

A note informs Jev. It suppresses nothing, and Jev still judges the code around it.
includeComments: true in the config sends every comment, for a project whose rules are about comments, such as doc comments on exports.

Suppressing a finding

A finding Jev got wrong is suppressed in the code, with a comment saying why:

// adhere-ignore alchemy/providers/idempotent-delete -- DeleteActivity succeeds on a missing activity; probed 2026-09-25
delete: Effect.fn(function* ({ output }) {
  yield* sfn.deleteActivity({ activityArn: output.activityArn });
}),

Comment	What it covers
`adhere-ignore`, on a line of its own	The statement that starts on the next line of code: above, the whole handler, wherever in it Jev points, since the lines it points at can move between runs.
`adhere-ignore`, at the end of a line of code	The statement that starts on that line.
`adhere-ignore-file`, anywhere in a file	The whole file. The rules it names are not judged there at all.

A comment names rules as a report does, separated by commas.
What follows -- is the reason, for whoever reads the code next.
A finding without lines is covered when its section overlaps the statement.
Jev never reads these comments, even with includeComments: true.
The summary counts the findings they suppress.

Cache

Judgments are cached in .adhere/cache/. Commit it. Every judgment is a paid request, and Jev's answers vary a little from run to run, so a committed cache gives everyone and CI the same findings without paying for them again, and a pull request changes the judgments of only the files it changes.

What changed	What is judged again
A file's code	Every rule, for that file.
A file's comments alone	Nothing.
A file is moved or copied	Nothing. Identical files share their judgments.
A rule's text, a matcher, or the source of its `appendState`	That rule, for every file.
What a rule `reads`, such as the workspace's packages	That rule, for every file.
The threshold is lowered	Nothing. Cached judgments newly above it are located.
adhere asks its questions differently, after an upgrade	Every rule, once.

To keep the cache out of diffs, mark it as generated in .gitattributes:

.adhere/cache/** linguist-generated -diff

Where lint gates a merge. Anyone who can commit can also write a cache file saying that code passes. Lint a pull request against the cache as merged, not as the pull request has it. Only the files it changes are judged again:
rm -rf .adhere/cache && git checkout origin/main -- .adhere/cache
adhere lint --yes

How the cache is stored and pruned

files/ holds Jev's answers by the content they are about. Each cache file is named for the hash of a source file's content as Jev reads it, without its comments.
Each answer sits under a fingerprint of the model and everything the rule sends: its text, its matchers, what it reads, and its appendState's source.
tallies/ keeps the linter check's answers, one tally per rule, under a key that changes with the rule's text and the model.
A cache file is written once and never changed: it is also named for the hash of its own text. Two branches that judge the same code add files rather than edit them, so git merges the cache without a conflict.
After a run that reads every file, adhere prunes the cache: it deletes answers about content no file has, and folds what is left into one file per content. A run narrowed by --filter does not prune.
Pruning keeps answers to rules the run left out, so a run with other presets or rules, or with one topic, loses nothing another run still asks. It keeps answers to a rule's old texts too, until the content they are about is gone.
It also keeps answers about a file the run read but judged against no rule, such as a test when no rule judges tests. A file the config excludes is not read, so answers about it go.

Agent skills

Three skills teach an agent to work with adhere:

Skill	Print it with	What it does
`adhere`	`adhere skill`	The reference: the commands, how to write and word a rule, the config, presets, shared rules, comments and suppressions, the cache, reading a report, and tuning a rule.
`adhere-setup`	`adhere skill setup`	Sets adhere up in a repo with you. See Generate rules from your repo.
`adhere-fix`	`adhere skill fix`	Verifies and fixes the findings `lint` reports: it checks each against its rule and the code, fixes the real ones, reports the false ones with the evidence, and says when a rule is wrong more often than right.

Install them with the skills CLI: skills add darkmatter/adhere. The binary carries all three, so adhere skill fix > .agents/skills/adhere-fix/SKILL.md works without a checkout.

In CI

Pipe the fix skill into an agent that runs without asking. The skill tells it that invoking the skill permits every adhere lint run and its cost, and adhere skill fix prints the skill of the adhere that runs, so its flags match:

adhere skill fix | codex exec --yolo
adhere skill fix | claude -p --dangerously-skip-permissions

Both turn off the agent's own permission checks, which suits a runner thrown away after the job. To keep Codex in its sandbox, let the sandbox reach api.typesafe.ai:

adhere skill fix | codex exec --sandbox workspace-write \
  -c sandbox_workspace_write.network_access=true -c approval_policy=never

Set TYPESAFE_API_KEY from a secret.
The agent adds no adhere-ignore comments or @adhere notes unless ADHERE_ALLOW_SUPPRESSIONS=1 is set in its environment, as in adhere skill fix | ADHERE_ALLOW_SUPPRESSIONS=1 codex exec --yolo. It then suppresses a finding it shows is false, or that three fixes did not clear, with the evidence in the comment. Without it, the agent leaves those findings, and its report ends by naming them and the variable.
End the job with adhere lint --yes. The agent exits 0 whatever it left, so this step passes or fails the job. It judges only the files the agent changed, and reads the rest from the cache.

Do not run it on pull requests from forks. The agent reads their code with your secrets in reach and the network open, so code written to steer it can send them out.

How a file is judged

Split. adhere splits the file into sections of whole statements, in runs of at most 30 lines.
Judge. One request per file asks Jev, for each rule, how likely the file is to break it.
Locate. A second request, only for rules above their threshold, asks which section and lines break the rule, and whether the file shows enough to decide.

A file too long for Jev's context is skipped and counted in the summary.

The steps in detail

Split. adhere splits each file into sections of whole statements, such as its imports, constants, and functions, packed in order into runs of at most 30 lines. A class or function longer than that splits into its methods or the statements of its body. Jev reads the file as these sections.

Judge: one request per file.

The state is the file's sections under their numbers from 1, as { "1": "…", "2": "…" }, without their comments (see Comments), and nothing else.
Each rule is one noul (yes/no probability) question that carries the rule's description and details, and its code under its words, must and never, or a guideline's should and should_not. It asks whether the file diverges from the pattern must shows, with never as an example of diverging, or, for a rule with only code that must never be written, whether the file contains that code.
Criteria for each answer draw the line at the rule's scope, so a file with no code the rule is about is a no.
No rule sits in the shared state, so a rule's probability depends only on the file and that rule, not on which other rules share the request. Every question shares the state cost of the request.
On up to 10 files per project rule, the rule's question has a second one beside it, the linter check (see Check for contradictions), with the same fields.
A rule with appendState or reads has requests of its own, here and below, whose state adds to this one.

Locate: a second request, only when at least one rule's probability is above its threshold. Per flagged rule:

When the file has more than one section, one choice question among the sections' numbers, which yields the section to report. The options carry no text of their own, since each is a key of the state: describing each section by its first line, or by what it declares, located no better on the eval.
For each section, two choice questions among its lines, each option the line's text: the line the violation starts on and the one it ends on. The chosen section's are underlined. On the eval, lines asked for in this request were named as well as lines asked for on their own, and they held a violation for 94% of real findings, most often as one line (study).
One noul carrying the rule: whether you can tell if code breaks it from code alone, without knowing what other files, libraries, services, or configuration do. It yields the context score and its warning.

Limits. Jev reads at most 32k tokens of state and one question together, and 64k tokens in a request. adhere estimates tokens from the JSON it sends, at about three bytes a token.

When a file's questions would not fit in one request, they are split across several.
A file whose code and longest question would not fit together, or that has more than 255 sections, is skipped and counted in the summary, and the rest of the run goes on.
So is a file Jev itself counts as over its context, which text denser than the estimate, such as CJK, can cause.

Upgrading

Names and layouts from older versions:

Version	What it had	What happens now
before 0.7	A rule's examples were `reference` and `avoid` in a config, and a fence could be tagged `avoid`.	They still read as `must` and `never`.
through 0.7	The `effect` preset checked seven conventions a linter checks exactly.	They are left to `@effect/tsgo`. The notes on `effect` under Presets say how to turn them on.
before 0.13	A rule was any `*.md` file in `.adhere/`, as `.adhere/data/brand-ports.md`.	A `*.md` file outside every rule's directory, but a `README.md`, refuses the run. The refusal says where each such file goes to keep its id, so `adhere-ignore` comments and `overrides` still name it.
before 0.13	`.adhere/tsconfig.json` included only `config.ts`.	Make its `include` `["*/.ts"]` to give rules written in TypeScript their types.
0.13.1	The workspace's packages had a key of their own, `includeWorkspacePackages: true`.	adhere no longer reads it, and a rule that still has it gets no packages. Write `reads: ["workspacePackages"]` in its place.
0.13.1 and before	Prose around a `RULE.md`'s code was ignored.	Jev reads it after the description, as the rule's details, so a rule with prose is judged again once. Keep prose about the rule: why it holds, or where it does not apply.
0.14.0 and before	The `effect` preset had `services/test-layers-are-in-memory` and `services/operations-have-no-requirements`.	They are gone, and an `overrides` entry naming either refuses the run: remove it. `@effect/tsgo`'s `leakingRequirements` checks the second. `adhere-ignore` comments naming them suppress nothing.

Development

From a checkout, with Bun:

bun install
bun link          # puts `adhere` on PATH
adhere lint --preset effect

In a checkout no platform package is installed, so bin/adhere.js runs src/main.ts with Bun instead.
bun run typecheck runs tsc and the linter.
bun run test runs the tests.

Native executable

Command	What it builds
`bun run build`	`dist/adhere`, a single binary with Bun and the presets inside it, through Bun's `--compile` with `--asset ./presets`. It runs without Bun or `node_modules` on the target machine and still loads the repo's `config.ts` and Markdown rules from disk.
`bun run build:npm`	The same for every published platform, each into its package under `dist/npm/`, with a copy for the GitHub release in `dist/release/` (`scripts/npm-packages.ts`).

Each platform's executable is its own npm package, @drkmttr/adhere-<platform>-<arch>, limited by os and cpu, and an optional dependency of @drkmttr/adhere, so an install fetches only its own machine's. The package's bin/adhere.js finds it and runs it with Node. The same executables are attached to each GitHub release, named for uname -s and uname -m as in adhere-Linux-x86_64, so installing one needs no Node.

adhere --version tells the builds apart:

Build	What it prints
A release	Its version, as `adhere v0.10.0`.
`bun run build`	`git describe --tags --dirty` where it was built, as `adhere v0.10.0-16-g9a41328`: 16 commits after v0.10.0, at `9a41328`, ending in `-dirty` when the tree had changes.
From source	The checkout's version, marked `(source)`.

Release

Releases are cut by CI from pushed version tags. From a clean, up-to-date main, push the release tag:

git tag v0.3.0
git push origin v0.3.0

The tag push starts .github/workflows/release.yaml, which checks out main, derives 0.3.0 from v0.3.0, and runs bun run release -- --ci 0.3.0.
Release-it bumps package.json, commits chore: release v0.3.0 back to main, skips npm publish, and creates the GitHub Release for the existing tag.
release.yaml then calls the reusable publish workflow, .github/workflows/publish.yaml, directly, avoiding a chained GitHub Release event created by GITHUB_TOKEN.
The publish workflow verifies that package.json matches the release tag, builds the platform packages, attaches their executables to the GitHub Release, and publishes each package before @drkmttr/adhere, which it first lists them in as optionalDependencies at the same version.

License

MIT. See LICENSE.

Name		Name	Last commit message	Last commit date
.adhere		.adhere
.github/workflows		.github/workflows
bin		bin
eval		eval
presets		presets
scripts		scripts
skills		skills
src		src
test		test
.gitignore		.gitignore
.prettierignore		.prettierignore
AGENTS.md		AGENTS.md
LICENSE		LICENSE
README.md		README.md
bun.lock		bun.lock
package.json		package.json
tsconfig.json		tsconfig.json

Latest commit

History

Folders and files

Repository files navigation