The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Skimless listing page.
English · Português
The reading order for pull requests that got too big.
Skimless turns a diff into a packet: what to read first, which files can wait, and which lines deserve a human. It runs on your machine. It does not call a model, post a comment, or phone home.
Zero runtime dependencies. Node.js 22 or newer. A CLI, a GitHub Action, a library, and an MCP server for AI agents. Packets in English and Portuguese.
The sample verify.ts calls timingSafeEqual on a buffer compared with itself. Skimless has no rule for that. It puts the file third, under the live-shaped key, so a person sees it. That is the product: a path through the diff, not a verdict.
English packet · Pacote em português · A real diff: Express 4.21.2 → 5.0.0
Pull requests written with an agent get big. The reviewer opens the first file, gets tired, and approves the lockfile with the same glance they would give a migration. Skimless tells you where skimming is safe and where it is not: the public API first, the sharp edges next, the noise at the end. The score is a reading priority. A quiet packet does not prove the diff is safe.
Or install it once with npm install -g skimless and call skimless.
--format all writes skimless.html, skimless.md, and skimless.json. Secret-shaped strings are redacted unless you pass --no-redact. --fail-on defaults to none locally. In CI, use high.
On a generated 2,000-file diff, a review packet comes back in about 100 ms on a laptop. npm run bench prints the number for your machine.
story is the default: API and types, then auth, signing, billing, migrations, and CI, then manifests, then the rest, then tests, docs, lockfiles, and generated code. A high finding jumps its file up into the sharp band. risk sorts by score alone.Copy examples/skimless.config.json to skimless.config.json to set the budget, ignore globs, and the CI failure level.
The action needs no token and posts no comment. The reading order and findings land in the job summary, on the pull request's Checks page. It passes inputs through the environment, then to git diff as arguments, never stitched into a shell string. docs/ci.md has the full workflow, every input, and a GitLab CI job.
skimless mcp is an MCP server. Claude Code, Cursor, VS Code, and other agents can ask for the reading order before they review a change, then read the files in that order.
Skimless still does not call a model. It serves one. Secret-shaped strings are always redacted in MCP responses, so a live key in the diff never reaches the agent's context. docs/mcp.md has the setup for each client, the tool arguments, and prompts that work.
Day to day, the CLI runs the TypeScript source directly. npm run build emits dist/, which is what npm ships.
Rules, reading order, and the library surface are described in docs/. The contributor notes are in CONTRIBUTING.md. The project is MIT, in GOVERNANCE.md terms: one maintainer at the start, decisions in the open, no pretended community.