Unpaired Unicode surrogates from tool results corrupt session history, causing all subsequent API calls to fail

Description

When a tool (particularly Playwright/browser automation) returns content containing unpaired Unicode surrogates, that data gets stored in the conversation history. All subsequent API requests for that session then fail with:

API Error: 400 {"type":"error","error":{"type":"invalidrequesterror","message":"The request body is not valid JSON: no low surrogate in string: line 1 column 222905 (char 222904)"}}

Once this happens, the session is permanently corrupted - every future request fails because the malformed history is included in each API call.

Steps to Reproduce

1. Start a Claude Code session
2. Use Playwright to navigate to a website with malformed UTF-16 encoding (e.g., sonnysfarm.com/products/one-dozen-eggs)
3. The tool result containing bad Unicode gets stored in history
4. All subsequent prompts in that session fail with the "no low surrogate" error

Expected Behavior

Tool results should be sanitized before being stored in conversation history. Unpaired surrogates (U+D800-U+DBFF not followed by U+DC00-U+DFFF, or vice versa) should be replaced with the Unicode replacement character (U+FFFD) or stripped.

Actual Behavior

Raw tool output is stored directly in history without sanitization, corrupting the session permanently.

Suggested Fix

Add Unicode sanitization when storing tool results in conversation history:

function sanitizeUnicode(str) {
  if (typeof str !== "string") return str;
  return str
    .replace(/[\uD800-\uDBFF](?![\uDC00-\uDFFF])/g, "\uFFFD")
    .replace(/(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/g, "\uFFFD");
}

Environment

• Claude Code CLI
• Observed when using Playwright MCP for web scraping
• The problematic characters typically come from websites with malformed HTML/encoding

Workaround

Currently the only workaround is to start a fresh session, losing all conversation context. Users cannot recover a corrupted session.

Impact

This is particularly disruptive for long-running sessions where significant context has been built up. One bad website scrape can destroy an entire session's worth of work.