KogenCredo

Credo checks for recurring generated-code mistakes and verification integrity: error handling, test quality, global state, structural limits and query regressions.

Version 0.1.0. Source: KogenAI/kogen_credo.

Start here

Check catalog: detectors, defaults, limits and source links.
Explicit configuration example: select checks and supply project-specific parameters before enabling them in CI.
Diff checks: compare changed functions with their base.
Migration guide: names moved from OptimumCredo.
Changelog: version history and release changes.

Installation

Add the Hex dependency:

{:kogen_credo, "~> 0.1.0", only: [:dev, :test], runtime: false}

Requires Credo 1.7.19 or newer in the 1.x series. When adopting this version, update Credo together with it; an older locked Credo does not provide the token metadata these checks need, and older releases cannot lint Elixir 1.20 reliably.

For development with a sibling checkout:

defp deps do
  [
    {:kogen_credo, path: "../kogen_credo", only: [:dev, :test], runtime: false}
  ]
end

Enable individual checks in .credo.exs:

%{
  configs: [
    %{
      name: "default",
      checks: %{
        enabled: [
          {KogenCredo.Check.Warning.SkippedTest, []},
          {KogenCredo.Check.Readability.JustifiedCoverageIgnore, []},
          {KogenCredo.Check.Warning.BroadRescue, []},
          {KogenCredo.Check.Warning.RepoAllThenTrim, []}
        ]
      }
    }
  ]
}

Run mix credo --strict to include low-priority advice. Neither installing this library nor copying a package example establishes the right policy for an application. Review actual findings and configure namespaces, codecs, domain roots, allowed serial tests and resource groups where needed.

OptimumCredo remains independent and owns code-review preferences. Enable both libraries where useful, choosing between overlapping policies deliberately.

Integrity and static-analysis limits

KogenCredo.Integrity provides suppression scanning and comparison with a golden Credo configuration. Parsing a configuration for comparison does not evaluate arbitrary Elixir. Store golden policy in a location the code-writing agent cannot change. The library cannot establish that trust boundary by itself.

NoCredoDisable is an advisory Credo detector, not a tamper-proof enforcement mechanism: Credo can suppress its own checks. The separate integrity scan exists for enforcement outside Credo's suppression pipeline.

mix kogen.unused_public reports public functions with no observed compiled caller. It is advisory by default and needs compiled debug information. Callbacks, entry points, external library consumers, dynamic dispatch and macros limit what static analysis can establish; read the task/module documentation before acting.

Module names

Analysed code names modules, it does not load them. Aliases, imports, defmodule nesting, __MODULE__ and module parameters are resolved to dotted names (MyApp.Repo; Erlang modules keep their colon, :ets) and compared as strings, so code whose modules are not compiled, not loaded or do not exist is analysed like any other, and no atom is created for a module the analysis only reads. Module parameters keep the form they always had: atoms for broad_exceptions, codec_modules, ForbiddenCall calls, stub_exceptions, shared and domain namespaces (a string is still rejected, so a typo cannot look like "nothing configured"); full-name strings for repo_modules. Repo names, function names and fields (including the diff CLI's --sensitive-field, --cast-sensitive-field and --repo-name) are compared by name and are never turned into atoms. Only mix kogen.unused_public handles real module atoms, because it reads installed BEAM debug information. KogenCredo.Integrity.parse_config/1 returns modules written in a config as KogenCredo.Integrity.ModuleRef structs for the same reason.

Framework boundaries and test scope

StringKeyAccess recognises literal Phoenix/Oban/Ecto callback boundaries, connection fields, JSON query paths and directly encoded wire documents. Other wire adapters still need codec_modules or codec_globs. Recognition is static; it does not track values across arbitrary calls or variable shadowing.

The ForbiddenCall async-test preset uses only_async: true rules and checks literal async: true modules. Explicit rules can inspect other scopes. SizeLimits excludes ~H markup from function length; use count_template_lines: true to count it.

Checks that infer test state through custom case templates need a case_templates map, for example %{"MyApp.DataCase" => :inherit, "MyAppWeb.ConnCase" => :inherit}. Supply it to the relevant checks before treating zero findings as coverage of those templates. Serial-case exemptions should describe reviewed shared resources.

Development

After mix deps.get, run make ci for formatting, compilation, tests, strict Credo and documentation checks. The individual commands are:

mix deps.get
mix test
mix format --check-formatted
mix compile --warnings-as-errors
mix credo --strict

The checked-in self-lint policy (mix credo --strict) reports no findings. Passing it, like test/format/compile success, says nothing about the precision of the detectors.

Test violations, valid lookalikes, parameter changes and lexical scope. Run the suite across supported Elixir versions before publication. New diff rules remain advisory until their precision is demonstrated on real history; unit tests alone are not that evidence.

Folder map: lib/kogen_credo/check/ owns Credo detectors; lib/kogen_credo/diff/ owns function comparison and reporting; lib/mix/tasks/ owns command entry points; remaining library modules own shared analysis helpers and integrity checks; test/ owns regression coverage; docs/ owns the catalog and usage/migration details.

Origin and license

EmptySetupBlock, JustifiedCoverageIgnore, RuntimeEnvDevDefault and SkippedTest originated in OptimumCredo and retain its copyright notice. Other checks grow from the retained Kogen research prototypes and original implementations. The two packages do not depend on one another.

MIT. See LICENSE.

Next Page → Changelog