[Feature]: ( CodeGen ) AI-powered test generation with network capture and prompt export

[BlazinArtemis]

🚀 Feature Request

Bridge playwright codegen with external AI tools by capturing the full recording context — user actions, network requests, and page state — and exporting it as a structured prompt that any AI assistant can use to generate production-ready Playwright tests.

Today, playwright codegen produces basic recorded code, but developers still spend significant time manually adding assertions, handling API responses, structuring test data cleanup, and following best practices. Meanwhile, AI tools like Claude Code, Cursor, and ChatGPT can generate excellent test files — if given the right context about what happened in the browser.

This feature would add an opt-in --ai-codegen flag that captures everything the recorder knows and makes it available to AI tools, without changing any default behavior.

Example

# Record with AI context capture enabled
npx playwright codegen --ai-codegen https://myapp.com

Workflow 1 — Copy/paste to any AI tool:

1. Run codegen with --ai-codegen, interact with the app
2. Click "Generate Test" — a structured prompt is copied to clipboard
3. Paste into Claude Code, Cursor, ChatGPT, or any AI tool
4. AI generates a complete test file with assertions, API mocking, and data cleanup

Workflow 2 — AI reads the session directly via MCP:

1. Run codegen with --ai-codegen, interact with the app
2. Tell your AI assistant: "there's a codegen session running, write the test"
3. AI reads the live session via a new recorder_get_session MCP tool
4. AI generates the test — no copy/paste needed

The prompt would include:

• Recorded user actions with timing
• Network requests classified by relevance (API calls vs. page loads vs. noise)
• Sensitive data automatically redacted (passwords, tokens, credit cards)
• Playwright best practices and data cleanup instructions

Motivation

playwright codegen captures what happened but not why it matters. The recorder knows about clicks, navigations, and form fills — but it doesn't capture the API calls those actions triggered, the response data that should be asserted on, or the test data that should be cleaned up. Developers bridge this gap manually today.

AI tools are already great at generating structured test code, but they lack context about the actual browser session. This feature closes that gap by making the recorder's full knowledge available as AI-consumable context.

Key benefits:

• Works with any AI tool — clipboard export means no vendor lock-in
• MCP integration — AI assistants that support MCP can read sessions directly, enabling a fully conversational workflow
• Zero impact on existing users — everything is gated behind --ai-codegen, default codegen behavior is completely unchanged
• No new dependencies — builds entirely on existing Playwright internals

[BlazinArtemis]

I have a working implementation of this at BlazinArtemis/playwright-codegen-pro if it's helpful for the discussion.

The branch includes:

• --ai-codegen flag wired through the full protocol stack (opt-in, zero impact on default behavior)
• Network capture & classification (networkCapture.ts)
• Sensitive data redaction (sessionRedactor.ts)
• Structured AI prompt builder (sessionPromptBuilder.ts)
• recorder_get_session MCP tool
• Live session file (.playwright-session.md) updated continuously via ThrottledFile
• 57 unit tests, all existing ctest / ttest suites pass
  Happy to open a draft PR if that's useful.

[dgozman]

@BlazinArtemis Thank you for the issue! Please take a look at playwright-cli that is aimed at coding agents and makes them capable to generate tests.

[BlazinArtemis]

Hi @dgozman — thanks for the pointer, I took a close look at playwright-cli and I think the two tools are solving fundamentally different problems.

playwright-cli is designed for AI agents that autonomously drive the browser — the AI opens pages, inspects snapshots, clicks, and generates code from its own exploration. This feature is the opposite: the human drives the browser naturally, and the system silently captures everything in parallel so an AI can generate a production-quality test from that real session context.

The network gap is also structural. playwright-cli network is an on-demand DevTools snapshot — useful for an agent inspecting what it just triggered. networkCapture.ts is a continuous capture layer that classifies requests by relevance (API call vs page load vs noise), correlates each to the UI action that triggered it, and feeds that into the AI prompt so assertions come from real backend responses, not guesses. sessionRedactor.ts then strips passwords, tokens, and card numbers before anything reaches the AI. None of that exists in playwright-cli. Their own test generation reference confirms it: "Generated code captures actions but not assertions. Add expectations manually." — that manual step is exactly what this feature eliminates.

	playwright-cli	--ai-codegen
Who drives the browser?	AI agent	Human developer
Continuous network capture?	No — on-demand snapshot	Yes — networkCapture.ts
Classified + correlated to actions?	No	Yes
Sensitive data redacted?	No	Yes — sessionRedactor.ts
Session exported for AI?	No	Yes — prompt + recorder_get_session MCP
Assertions from real responses?	No — added manually	Yes
Extends codegen?	No — replaces it	Yes

[dgozman]

@BlazinArtemis Thank you for the explanation. I don't think we are interested to explore this particular direction right now. I'll leave the issue open for prioritization, but don't expect any movement in the nearest future.