Open-source statement-level Playwright tracer, purpose-built for AI agents.

SaaS | Website | Docs

@heal-dev/heal-playwright-tracer

heal-playwright-tracer is an agent-first diagnostic layer for your Playwright tests. It gives agents (and humans) everything they need to quickly analyze test results.

👉 Add this to your playwright config, run your tests, point Claude to the improved heal trace, get more accurate test diagnosis

Why

The playwright trace doesn't contain enough data for LLM-based agents such as Claude or Open Code to analyze tests results reliably. That's because the trace ifs focused on locator evaluation, while real-life tests also evaluate non-playwright code. Heal adds the missing instrumentation layer to let LLM agents work their magic. And it's useful for humans in complex test codebases, too!

Feature	Playwright Trace	Heal Tracer	Example: What Heal Adds
Granularity	Action-level	Statement-level	Shows let x = calculate() line-by-line, not just the final page.click().
Data Format	ZIP/Binary	NDJSON Stream	{"type":"step","file":"auth.spec.ts","line":12,"val":{"user":"dev"}}
Visual Context	Standard screenshots	Highlighted locators	An image where the target button is outlined in a neon overlay to prove hit-box accuracy.
Variable State	Limited/Debugger only	Full Variable Values	Captures that status_code was 403 inside a hidden helper function.
Error Detail	Standard stack trace	Serialized Errors	A JSON object containing the DOM snapshot at the exact millisecond of the throw.
Timing	Action durations	Per-statement timing	Identifies that a specific if statement logic took 2.5s to evaluate.
Correlations	Loose logs/network	API Correlations	Links Trace_ID_99 directly to Source_Line_45 in the NDJSON stream.

Install

npm install -D @heal-dev/heal-playwright-tracer

Wire the Babel plugin and the reporter in playwright.config.ts:

// playwright.config.ts
import { defineConfig } from '@playwright/test';

export default defineConfig({
  // @ts-ignore — `babelPlugins` is a supported Playwright option not yet in its public types
  '@playwright/test': {
    babelPlugins: [
      [
        require.resolve('@heal-dev/heal-playwright-tracer/code-hook-injector'),
        { include: [/\/tests\//] },
      ],
    ],
  },
  reporter: [['@heal-dev/heal-playwright-tracer/reporter']],
});

Both are required: if you wire the Babel plugin without the reporter, the fixture fails fast on the first test of every worker. For why the reporter is mandatory (crash rescue, Playwright attachment copy, execution-history index), see docs/configuration.md.

If you prefer to keep the config fully typed, declare the babelPlugins option once at the top of the file instead of using @ts-ignore:

declare module '@playwright/test' {
  interface Config {
    '@playwright/test'?: {
      babelPlugins?: Array<[string, object?]>;
    };
  }
}

Per-test output lands at heal-traces/<executionId>/<playwrightTestId>/<attempt>/heal-traces.ndjson. See Output layout below for the full tree.

Output layout

Every run produces a self-contained execution dir under: heal-traces/<executionId>/<playwrightTestId>/<attempt>/heal-traces.ndjson.

<cwd>/heal-traces/
├── executions.ndjson                                   # append-only run index
└── <executionId>/                                      # auto-generated uuidv4 per run
    ├── execution.json                                  # per-run manifest
    └── <playwrightTestId>/                             # Playwright testInfo.testId
        └── <attempt>/                                  # 1-indexed
            ├── heal-traces.ndjson
            ├── trace.zip                               # if Playwright recorded one
            ├── screenshots/
            │   └── stmt-0001.png
            └── videos/
                └── video.webm

Keys are shaped {executionId}/{testId}/{attempt}/... so the layout maps cleanly onto any object-store key prefix if you ship traces elsewhere later.

History is append-only — runs accumulate, nothing is pruned. Clean up manually with rm -rf heal-traces/<id>/ when needed.

Note: add heal-traces/ to your .gitignore — these are local build artefacts and should never be committed.

<executionId> is a uuidv4 the reporter generates at the start of each run; the trace viewer uses it to scope and group artefacts per run.

Usage

1. After installing Heal, run your tests with the usual npx playwright test command.
2. You should see heal-traces.ndjson.
3. You can ask Claude or another agent to use those to understand your test results.

Viewing traces

A small CLI ships alongside the tracer so humans can browse the captured traces in a local browser without setting up a Heal account:

npx heal-tracer view

The CLI starts a local HTTP server on a free OS-assigned port, reads <cwd>/heal-traces/, lists every recorded execution in the header dropdown (newest first), and opens the latest one in your default browser. Switch executions from the dropdown to inspect history. Press Ctrl+C to stop.

Options

Flag	Description
--port <port>	Bind a fixed port instead of an OS-assigned ephemeral one. Useful when an external frontend needs a stable URL. Falls back to the HEAL_TRACER_PORT env var.
--api-only	Serve only the /api/* routes — no SPA bundle, no browser auto-open. For driving the REST API from a separate frontend dev server.
--verbose	Mirror spawned subprocess output (heal analyze, heal login) to the tracer console.

To consume the REST API from your own frontend during development, run the server on a fixed port without the bundled UI:

npx heal-tracer view --api-only --port 3001
# or, via env (e.g. from a .env loaded with `node --env-file`, direnv, or your shell):
HEAL_TRACER_PORT=3001 npx heal-tracer view --api-only

CORS is wide-open (Access-Control-Allow-Origin: *), so a frontend on its own dev server can fetch('http://localhost:3001/api/executions') cross-origin with no extra setup. The API is unauthenticated and localhost-bound by design — don't expose it on a public interface.

Claude Skill

See this Claude skill for a starter.

Sample output

heal-data/heal-traces.ndjson — one record per line:

{"kind":"test-header","schemaVersion":1,"test":{"title":"it works","file":"tests/example.spec.ts","context":{"testId":"...","attempt":1}}}
{"kind":"statement","statement":{"loc":{"line":5},"source":"await page.goto('https://example.com')","durationMs":412,"status":"ok","children":[...]}}
{"kind":"statement","statement":{"loc":{"line":6},"source":"await expect(page.getByRole('heading')).toBeVisible()","durationMs":73,"status":"ok"}}
{"kind":"test-result","status":"passed","duration":1234,"stdout":"...","stderr":""}

Schema: src/domain/trace-event-recorder/model/statement-trace-schema.ts (also exported as @heal-dev/heal-playwright-tracer/statement-trace-schema).

Screenshots

Every statement that calls a patched Playwright locator action (click, fill, hover, press, …) or a locator assertion (expect(locator).toBeVisible(), toHaveText(), …) produces a PNG screenshot with the targeted element outlined via an overlay drawn in-page — so the agent sees what Playwright was actually pointing at at the moment the action ran, not just the raw page.

Files are written to the per-test heal-data/ directory and referenced on the corresponding statement via the screenshot field:

{"kind":"statement","statement":{"source":"await page.getByRole('button', { name: 'Submit' }).click()","status":"ok","screenshot":"stmt-0007.png"}}
{"kind":"statement","statement":{"source":"await expect(page.getByRole('alert')).toBeVisible()","status":"ok","screenshot":"stmt-0008.png"}}

Statements that don't touch a locator (plain JS, utility calls, page.goto) have no screenshot field — capture is scoped to the Playwright surface where it adds diagnostic signal.

Architecture, and extending the tracer

See development.md

Caveats

The Babel plugin rewrites every leaf statement with a try/catch/finally and three hook calls — the same shape of transformation Istanbul applies for code coverage. Two consequences to be aware of:

• Instrumented files are larger. Each statement gains a wrapper, so on-disk size of transformed test files grows noticeably (typically ~2–4×, depending on statement density). This affects the files Playwright loads into workers, not your application bundle.
• Tests run slightly slower. The per-statement hook overhead is small in absolute terms but not free — expect a modest slowdown on CPU-bound test code. I/O-bound tests (the common case: await page.click(...), network, navigation) are dominated by the browser and barely move.

Scope the include filter in playwright.config.ts so only your tests/ directory is instrumented — never your app code or node_modules — to keep the cost contained.

License

Copyright © 2026 MYIA SAS.

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0), except for the tracer-viewer-bundle which is a vendored, free to use version of the Heal trace viewer. See the LICENSE file for the full text.