fix: filter base64 image from screenshot result

[Burry]

Changes

• Filter out base64 image data from the screenshot result to reduce context usage.
• Extract file paths from text content and attach guidance.

---

See #104

I have emailed [email protected] with a signed CLA.

[gaojude]

Thanks for tackling this issue! However, I discovered that playwright-mcp actually has a built-in flag for this use case.

Better Solution: --image-responses Flag

Instead of filtering the response after the fact, we can prevent the base64 image data from being sent in the first place by using the --image-responses flag when starting the playwright-mcp server:

// In src/_internal/browser-eval-manager.ts:66
const args: string[] = [
  "@playwright/mcp@latest",
  "--image-responses", "omit"  // Add this
]

Why This Approach is Better

1. Cleaner - Prevents image data at the source rather than filtering after transmission
2. More efficient - Never transmits base64 data over the stdio transport in the first place
3. Simpler - No need for filtering logic, regex parsing, or unit tests
4. More reliable - Doesn't depend on parsing text formats that might change

The screenshot will still be saved to disk (as your tests confirm), and the response will include the file path in the text - we just won't get the massive base64 blob consuming context.

Reference

The --image-responses flag accepts:

• "allow" (default) - sends base64 image data
• "omit" - excludes base64 image data from responses
• "auto" - sends images only if client can display them

See the playwright-mcp documentation for more details.

Would you be interested in updating this PR to use the flag approach instead?

[Burry]

@gaojude Thanks for the review. Good find—that's much cleaner! I just pushed a change to use that flag.