fix(mcp): browser-use screenshot saving path and inline rendering

[Gucc111]

Summary

• Set cwd: outDir on the browser-use MCP spec so that screenshots (both auto-named and user-specified filename) are saved to projectRoot/.pilotdeck/browser_screenshots/<sessionId>/ instead of an unpredictable inherited CWD.
• Add a file-reading fallback in marshalMcpContent(): when @playwright/mcp returns only a text block with a Markdown image link (no base64 — happens when the LLM passes filename), read the referenced file from disk and inject it as an inline image block, so the existing toolResultImages rendering pipeline can display it.

Root cause

@playwright/mcp has two path-resolution strategies for browser_take_screenshot:

LLM passes filename?	File saved to	Response contains
No	--output-dir (absolute)	text + image (base64) blocks
Yes	workspaceFile(filename) → relative to process CWD	text block only (Markdown link)

PilotDeck spawned the browser-use MCP process without setting cwd, so process.cwd() was inherited from the PilotDeck parent — an unpredictable location. Screenshots with filename were written to the wrong directory (or silently failed), and even when the file existed, the response lacked base64 data for the UI to render.

Test plan

• Start a session, ask the agent to take a screenshot (triggers browser_take_screenshot)
• Verify the screenshot file appears in projectRoot/.pilotdeck/browser_screenshots/<sessionId>/
• Verify the screenshot renders inline in the tool result area of the Web UI
• Refresh the page and verify the screenshot is still visible (session history path)
• Test with a model that does NOT pass filename — verify base64 inline rendering still works
• Test with a model that passes filename — verify the file-reading fallback kicks in

Made with Cursor