@simonlunay/redline-agent

The AI fix loop for Redline, "ESLint for designs". It checks a design, lets Claude repair the problems, re-checks, and repeats until the design is good. It returns the best version plus a full, replayable history.

$ npx @simonlunay/redline fix poster.json --out fixed.json --render-steps steps/

 redline fix  poster.json  target 90 · max 4 · claude-sonnet-5-5 (medium)

  #0  start          6  █░░░░░░░░░       11 errors · 3 warnings
  #1  accepted      94  █████████░  +88  0 errors · 2 warnings
        • move headline by (+0, -130) — move the headline above the product
        • set headline color to #0f172a — contrast on the sky
        … 19 more

  ✔ 6 → 94 in 1 iteration · errors 11 → 0 · stop: target reached

Install

npm install @simonlunay/redline @simonlunay/redline-agent

Installing this package enables the redline fix command of @simonlunay/redline. Set ANTHROPIC_API_KEY in the environment or in a .env file. Use --editor suggested to run offline: it applies the checker's own suggested fixes and needs no key.

CLI

| Option | Description | | ---------------------- | ---------------------------------------------------------------------- | | --out <path> | Where to write the best design (default <name>.fixed.json) | | --target <score> | Done when the score reaches this and no errors remain (default 90) | | --max-iterations <n> | Max repair iterations (default 4) | | --model <id> | Claude model (default claude-sonnet-5-5) | | --effort <level> | low medium high xhigh max (default medium) | | --editor <name> | anthropic (default) or suggested (offline, deterministic) | | --render-steps <dir> | Annotated PNG per iteration | | --no-vision | Don't send the annotated render to the model | | --config, --format | Checker config file; pretty or json output |

The exit code is 0 if the final design has no errors, 1 if errors remain, and 2 for usage errors.

Library

import { createAnthropicEditor, runFixLoop } from '@simonlunay/redline-agent';
import { createFixSession } from '@simonlunay/redline-agent/node';

const session = await createFixSession('poster.json');
const result = await runFixLoop(session.loaded.design, {
  editor: createAnthropicEditor({ model: 'claude-sonnet-5-5', effort: 'medium' }),
  check: session.check,
  renderImage: session.renderImage, // optional: let the model see the render
  target: 90,
  maxIterations: 4,
});

result.best.design; // the best design seen
result.history; // every iteration: design snapshot, report, edits + reasons, usage
result.stopReason; // 'target-reached' | 'no-issues' | 'max-iterations' | 'no-improvement' | 'editor-error'

The main entry point is isomorphic. ./node adds file loading, rendering, and the CLI command.

How the loop works

Each iteration: check → the editor proposes edits → applyFixes → guardrails → re-check.

• The model receives: the score, every issue with the checker's suggested fix, a layout table, feedback from failed attempts, the design JSON, and the annotated render.
• The model answers through a strict submit_edits tool. The edit format is the same one applyFixes uses: move, resize, setColor, setFontSize, setFontWeight, insertShape, each with a short reason.
• Layout and style only: no op can change text content, font family or image src, and a backstop compares those fields against the original after every step.
• Validation: invalid responses are retried once with the validation error.
• Rollback: an iteration is kept only if the score doesn't drop. Otherwise it's rolled back, and the next request explains what it broke.
• Stopping: the loop stops when score ≥ target with no errors, no issues remain, max iterations is reached, two iterations in a row don't improve, or the editor fails.
• Reproducibility: no server-side model fallback is enabled, so every response comes from the requested model.

DesignEditor is a one-method interface, so other LLM providers can plug in. createSuggestedFixesEditor() is the deterministic editor used for tests and as a rules-only baseline.

Evaluation

Mean score on 14 fixtures, 3 runs per model: 55.1 before, 80.0 with rules alone, 98.8 with Sonnet 5.5, 99.1 with Opus 5.5. Both models cleared every error in every run. See the main README for the full table, the method and the caveats.

License

MIT