notignored

Find every lint and type-check suppression comment in a codebase — natively, and fast.

$ notignored src/
src/app.py:3:12 ruff F401 (line) -- re-exported for the public API
src/app.py:5:58 ruff E501 (line) -- long wrapped URL
src/app.py:10:17 ruff * (line)
src/vendored.py:1:1 ruff E501 (file) -- vendored upstream, not ours to reformat
notignored: 4 ignores in 2 files

Why

Suppression comments are where lint and type-check debt hides. A # noqa costs one line to add and scrolls past review unremarked — especially in agent-written code, where silencing a rule is often easier than fixing it.

notignored turns every suppression into a first-class, queryable record: which tool, which rules, the stated reason, and exactly where it lives. That makes a high-level review of a large change possible — you can see the bypasses and their justifications without reading every line.

It parses the directives itself and never invokes the tool whose rule is being silenced, so scanning a tree costs a read and a scan instead of ten linter startups. That is what makes it cheap enough to run on every pull request.

Install

# From PyPI or npm — both ship the prebuilt binary, so no Rust toolchain is
# needed and nothing is compiled at install time:
pip install notignored-cli
npm install -g notignored-cli

# Or without installing at all:
npx notignored-cli src/

# Cross-platform, from source (Linux, macOS, Windows):
cargo install --git https://github.com/nickderobertis/notignored --locked

# Or a prebuilt binary (Linux, macOS, and Windows under a POSIX shell):
curl -fsSL https://raw.githubusercontent.com/nickderobertis/notignored/main/scripts/install.sh | sh

All four install the same notignored command. The notignored-cli distributions carry the release binary for your platform — a wheel per platform on PyPI, a package per platform on npm, picked automatically — so they are the fastest path on a CI image with no Rust toolchain, and the only one that works where github.com is blocked but the package registries are not. Prebuilt targets are Linux (x64, arm64), macOS (x64, arm64), and Windows (x64); anywhere else, cargo install builds from source.

The installer honours NOTIGNORED_VERSION / NOTIGNORED_INSTALL_DIR (or the --version / --to flags), verifies the archive against the SHA-256 checksum published beside it, and refuses to install a binary it cannot verify. Every tagged release attaches per-platform archives built on native runners.

In CI there is nothing to install: the GitHub Action fetches the release binary itself and posts what the pull request added.

- uses: nickderobertis/notignored@main

Try it

examples/ holds a handful of tiny files — Python, TypeScript, shell, Rust — each carrying the kind of suppression, and the kind of reason, real code collects. Point the binary at them:

$ notignored examples/
examples/api_client.py:3:20 mypy import-untyped (line) -- the vendored SDK ships no type stubs
examples/api_client.py:4:28 ruff F401 (line) -- re-exported so callers can configure retries
examples/deploy.sh:8:1 shellcheck SC2086 (next-line) -- the flags file is ours, and has to split into separate arguments
examples/deploy.sh:13:3 llmlint tool_output_is_signal (file) -- example input the README quickstart scans, not a script this project runs
examples/retry.rs:6:1 rust dead_code (next-line) -- the scheduler starts calling this once backoff lands
examples/widget.ts:6:3 eslint no-console (next-line) -- the mount path is traced in production
examples/widget.ts:9:3 typescript * (next-line) -- the vendored analytics global is declared without its options bag
notignored: 7 ignores in 4 files

Seven suppressions, seven tools, one pass over four files — no linter was run. Each line is path:line:column tool rules (scope) -- reason; --format json gives the same records as the envelope below, and --format markdown gives the comment the action posts. That block is checked against the real binary by tests/e2e/examples.rs, so it is output, not an illustration.

Usage

notignored [PATHS...] [--format human|json|markdown] [--tool NAME]... [--fail-if-found]
           [--diff [--diff-base REF]] [--github-repo OWNER/REPO] [--github-sha SHA]

• PATHS — files and/or directories. Directories are walked recursively, honouring .gitignore. Defaults to ..
• --format — human (default), json, or markdown (a pull-request comment body; see the action below).
• --tool — only report this tool; repeat to allow several. Omit for all.
• --fail-if-found — exit 1 when any suppression is reported.
• --diff — report only the suppressions the change added (see below).
• --diff-base — the git revision or range --diff compares against.
• --github-repo / --github-sha — the owner/repo and commit the markdown format builds its permalinks from.

Reviewing a pull request

A reviewer cares about the suppressions a change introduces, not the inventory it inherited. --diff reports only those: a directive is new when the diff added at least one of the lines it occupies.

$ notignored --diff --diff-base main --fail-if-found
src/app.py:42:20 ruff E501 (line) -- long wrapped URL
notignored: 1 ignore in 1 file

• Bare --diff compares the work tree — staged and unstaged — against HEAD.
• --diff-base REF takes any git revision or range. A plain ref is compared from the merge base, the way a pull request's "Files changed" is, so commits that landed on the base branch after this one forked are never reported as this branch's own. An explicit A..B range is passed to git as-is, two-dot semantics and all. These are llmlint's --diff / --diff-base semantics exactly.
• PATHS still narrow the result: notignored --diff --diff-base main src/ reports the new suppressions under src/ only. An empty intersection is a clean exit 0.
• Only the files the change touched are read, so a diff run stays fast on a large repository. Files the change deleted are skipped; a renamed file reports what the change added to it, not the lines that merely moved.
• Git names a file in bytes and a report names it with a string, so a path that is not valid UTF-8 has no faithful spelling here. It becomes an errors entry (and exit 2) rather than a file quietly dropped from the review — the lossy spelling would name a file that does not exist.

--diff shells out to git — infrastructure, not one of the linters whose directives are parsed natively — so it needs git on PATH and a work tree.

On a pull request: the GitHub Action

The action posts one sticky comment naming every suppression the pull request added, with its stated reason and a link to the line. It edits that same comment on each push instead of adding another, and on a pull request that adds none it posts nothing at all.

# .github/workflows/notignored.yml
name: notignored

on:
  pull_request:

permissions:
  contents: read
  pull-requests: write

jobs:
  suppressions:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0 # the base branch has to be fetched to diff against it
      - uses: nickderobertis/notignored@main

It exposes count (how many suppressions the change added) and report-path (the JSON report), so a later step can fail the build, upload the report, or gate on a threshold:

      - uses: nickderobertis/notignored@main
        id: notignored
      - if: steps.notignored.outputs.count != '0'
        run: echo "this change adds ${{ steps.notignored.outputs.count }} suppression(s)"

The steps are bash, so the action runs on the Linux and macOS runners; it needs jq and the gh CLI (both preinstalled there), plus cargo when version: local.

--format markdown renders exactly the body the action posts, so it can be previewed locally:

$ notignored --diff --diff-base main --format markdown \
    --github-repo nickderobertis/notignored --github-sha "$(git rev-parse HEAD)"

Both permalink flags are optional; without them each location renders as plain path:line text. When a body names fewer than four suppressions, each one also carries the source line with two lines of context on either side.

Exit codes

Findings go to stdout; the summary and any errors go to stderr, so notignored --format json > report.json captures only the report. A downstream consumer that stops reading (| head, | grep -q) is not an error: the scan's own verdict still decides the exit code.

Output

The json format emits the full report envelope:

{
  "version": 1,
  "ignores": [
    {
      "tool": "ruff",
      "scope": "line",
      "rules": ["E501"],
      "reason": "long wrapped URL",
      "path": "src/app.py",
      "line": 12,
      "end_line": 12,
      "column": 20,
      "raw": "# noqa: E501  # long wrapped URL",
      "suppressed": { "start_line": 12, "end_line": 12 }
    }
  ],
  "errors": []
}

• scope — line, next-line, file, or block.
• rules — rule names/codes exactly as written. [] means a blanket suppression of every rule the tool would apply (rendered as * in the human format).
• reason — the stated justification, or null. Taken from the tool's native reason syntax where one exists, otherwise from the trailing comment on the directive line; whitespace is collapsed to single spaces.
• line / end_line / column — 1-based.
• suppressed — the best-effort range the directive silences. end_line is null when it runs to end-of-file or is unterminated.
• errors — files that could not be read. Never a panic.

version is the envelope version; it changes only when the shape does.

Supported tools

Scope follows each tool's own rules, not a house convention:

• rust — an outer attribute is next-line and its suppressed range runs through the end of the item it annotates; an inner #![…] is file. A reason = "…" is the record's reason, even when the string wraps.
• shellcheck — a directive above the first command is file; anywhere else it is next-line. A directive ShellCheck itself rejects (trailing prose with no #, or one placed after a command) is reported by neither tool.
• typescript — the parity claim is pinned to one compiler: the typescript version in tests/js-toolchain/package.json — 7.0.2, the Go port — which is what tests/e2e/typescript_parity.rs drives. The 5.x compiler is a separate implementation of the same directives and is not guaranteed to read every form the same way, so the claim here is parity with the pinned compiler rather than with every tsc ever shipped.
• llmlint — ignore is line, ignore-file is file, and ignore-block … ignore-end is one block record spanning both directives. A block left unclosed keeps a null suppressed.end_line and adds an errors entry.

Each tool's own reason syntax is what gets captured: ESLint's -- description, Biome's mandatory : explanation, ruff's trailing # comment, and — for TypeScript, which defines no separator — whatever text trails the directive. A directive that lists no rules is a blanket suppression (rules: []).

Scope follows the tool rather than the syntax. // eslint-disable-line is line; // eslint-disable-next-line, // biome-ignore and // @ts-expect-error are next-line; # ruff: noqa, // biome-ignore-all and // @ts-nocheck are file; and the delimited pairs (/* eslint-disable */ … /* eslint-enable */, // biome-ignore-start … // biome-ignore-end) are block, running to end-of-file with suppressed.end_line: null when they are never closed.

Adding one is four touch points: a module under src/tools/, one line in src/tools/mod.rs::registry(), one row above, and a directive in tests/fixtures/polyglot/. tests/tools_contract.rs fails the build if a row here and the registered parsers disagree, and tests/e2e/polyglot.rs fails it if a registered tool is missing from that fixture tree.

Where a directive reaches, and who honours it

Every # type: ignore and # pyright: ignore is line-scoped; the module-wide forms are mypy's two # mypy: config comments and pyright's rule override; for ty, where the comment sits is the scope; and a Rust attribute reaches to the end of the item it annotates:

A Rust attribute's scope is next-line — that is where the item it annotates starts, and where a reviewer has to look — while its suppressed range covers the whole item, however many lines that item runs to. An inner #![…] attribute exempts the file it opens.

Pyright's <rule>=<value> override is reported only for the two values that switch a rule off, false and none; true, error, warning, and information turn a rule on or move its severity, so they are configuration rather than suppression. Pyright reads the rest of that line as its own item list, so the form can carry no reason — and a comment it refuses (a trailing # why, a value outside those six, or a directive that does not open the comment) silences nothing and is not reported.

Several tools honour a directive they did not invent: pyright and ty both act on mypy's # type: ignore, and ruff, pyright, and ty all act on one that does not open its comment (mypy does not). A directive is reported once, under the tool whose syntax it is — # type: ignore is one mypy record, not three — so a count of records is a count of suppressions written, not of checkers affected.

One line can still carry directives for several tools, and each record covers its own directive only. Given

import legacy  # type: ignore[import-not-found]  # no stubs published  # noqa: F401  # imported for its side effects

the mypy record's raw stops at # no stubs published and its reason is "no stubs published"; the ruff record's raw starts at # noqa: F401 and its reason is "imported for its side effects". A record's raw and reason always end where the next tool's directive begins, so one tool's live suppression can never be filed as another's justification.

# pyright: basic and # pyright: strict switch pyright's type-checking mode rather than silencing a diagnostic, and are deliberately not reported.

How it works

Source is scanned once per file by a language-aware comment extractor (src/comments.rs) that understands # comments, // line comments, multi-line /* … */ blocks (nested, for Rust), Rust attributes, and the punctuation that delimits a Rust item — and that knows a string literal when it sees one, so MESSAGE = "# noqa: E501" is never reported. Tool parsers consume that extraction; they never re-scan raw lines.

Development

just bootstrap   # from a clean clone
just check       # the full gate: format, clippy, tests + coverage, docs
just --list      # everything else

This is an Nx monorepo of three projects — the notignored crate at the repo root, plus the scaffolded python/notignored-sdk and npm/notignored-sdk SDKs. The repo-wide verbs above fan out across all of them; just nx run notignored-sdk-python:check runs one project's gate alone, and just nx show projects lists the graph. Pull-request CI narrows the gate to the projects the diff can reach.

just check runs the end-to-end suite, which drives the compiled binary as a subprocess and the real, pinned tools — ruff, mypy, pyright, ty, shellcheck, and llmlint (see the .<tool>-version files), eslint / biome / tsc (see tests/js-toolchain/package.json), plus the pinned toolchain's own rustc and clippy-driver — to prove that what notignored reports is what those tools actually suppress. just bootstrap installs them all under .dev/; it needs uv and Node.js 20+ on PATH.

License

MIT — see LICENSE.