RFC: reboot as a minimal Playwright test-generation eval—or archive

[BaseInfinity]

Context

PlaywrightGPT was useful in 2023 as a demonstration of natural-language browser control and Playwright test generation. The ecosystem now provides maintained versions of most of that surface:

• Microsoft Playwright MCP: https://github.com/microsoft/playwright-mcp
• Playwright Test Agents (planner, generator, healer): https://playwright.dev/docs/test-agents
• Playwright codegen: https://playwright.dev/docs/codegen

The current repository also carries substantial legacy weight: an early LangChain/AutoGPT stack, HNSW memory, custom HTML parsing, dynamic execution of model-generated JavaScript, stale model defaults, and large demo media.

A dependency-only refresh would produce another generic browser agent without a clear reason to exist.

Proposed decision

Run a small spike to decide between:

1. Reboot: make this a tiny, provider-neutral evaluation harness for natural-language-to-Playwright-test generation.
2. Archive: preserve it as a historical project, add a maintenance-status banner, and point users to current first-party alternatives.

Do not begin a broad modernization until the spike demonstrates a distinct use case.

Possible differentiated scope

The reboot should evaluate generated test artifacts, not compete as another general-purpose browser controller:

• provide a small set of local deterministic web fixtures and natural-language tasks;
• accept generated Playwright specs from pluggable model/agent adapters;
• run the specs in isolated Playwright contexts;
• report pass/fail, repair attempts, locator quality, runtime, and optional token/cost metadata;
• preserve generated specs, traces, and structured results for comparison;
• never execute arbitrary generated code implicitly—generation and --run should be separate, explicit steps.

Lightweight constraints

• Prefer @playwright/test and standard-library code.
• Remove LangChain, AutoGPT, vector memory, JSDOM/custom DOM serialization, and native HNSW dependencies unless the spike proves one is essential.
• Target Node LTS and TypeScript.
• Aim for no more than five runtime dependencies and roughly 500 lines for the initial core.
• Keep provider integrations behind a small adapter boundary; one reference adapter is enough initially.
• Add one deterministic fixture, one generation task, one scoring report, and CI.
• Do not rewrite git history to remove the existing demo assets.

Decision gates

Proceed with a reboot only if the spike proves all of the following:

• it measures something that Playwright MCP/Test Agents do not already expose as a reusable comparison artifact;
• a new contributor can install and run the example in under five minutes;
• the generated test is executable and its result is scored deterministically;
• the core stays within the lightweight constraints above;
• ongoing maintenance is bounded and does not require chasing every frontier-model SDK.

If those gates fail, archive the repository and update the README with historical context and links to maintained alternatives.