Skills that save your agent’s context for the real work.

Keep the full log out of contextsystem-one-verify · System One Skills

Run a noisy test or build once, hand your agent a short result, and keep the full log on disk. Across 563 real runs it returned 35% less text. It needs neither Sys1 nor a model.

Use compact check output →

Review the changesys1-review

Check a batch of Git changes against your repository’s rules. The preview is free, an unchanged batch costs no new requests for 24 hours, and findings you have already seen stay out of the output. Turn a real mistake into a new rule. Two bundled rules cover empty catch blocks and removed test assertions in JavaScript and TypeScript.

Set up agent review →

Check the completion claimsys1-verify

Before your agent says “done,” compare its message with Git, pull requests, and live pages. A claimed push with nothing pushed shows up as a contradiction. Missing evidence stays unverifiable. Works with a message from any coding agent.

Verify a completion claim →

Review scores rank candidates; they are not calibrated defect probabilities. Keep your repository’s tests and review.

We will publish results, including the ones where another pack wins, with the task set and raw logs. Until then, treat any comparison you read here as a question, not a result.

Sys1 already learns on a small scale: draft a rule from a mistake your agent made, and every later review checks for it. Sys1 and ALGAL are separate open-source projects today that call the same Jev decision API. We have measured cost per decision and output size, not whole-task token savings yet.

1. Supply contextSend state and the questions your application needs answered.
2. Choose a backendPin Jev, an experimental local model, or a compatible server.
3. Validate the responseSys1 checks the answer format and reports malformed responses as errors.
4. Act in your codeYou decide the thresholds, permissions, and next action.

import { createClient } from "@hraness/sys1/client";

const sys1 = createClient();
const { response, metadata } = await sys1.evaluate({
  state: "The build failed after a dependency update.",
  questions: {
    next: {
      type: "choice",
      criteria: {
        repair: "Fix the build",
        continue: "Continue work"
      }
    }
  }
});

console.log(response.answers.next);
console.log(metadata.backend);

The client also works in Node without the optional native runtime. The CLI and embedded router require Bun. Choose an integration or inspect the evaluation records.

npm install --global --allow-scripts=node-llama-cpp \
  https://github.com/hraness/sys1/releases/download/v0.17.0/hraness-sys1-0.17.0.tgz
sys1 doctor
sys1 review setup codex

The CLI requires Bun 1.3.14 or newer on macOS, Linux, or Windows. --allow-scripts=node-llama-cpp permits the native inference package’s install script. Installation downloads no model weights and activates no hosted backend. Use setup claude-code or setup devin for those agents.