Description

When the Anthropic streaming API sends SSE data: lines containing unescaped control characters (C0 range U+0000–U+001F), the @anthropic-ai/sdk streaming parser throws:

Bad control character in string literal in JSON at position N (line 1 column N+1)

This error is surfaced verbatim to the user as a message in the chat, with no retry or graceful recovery. The response is lost and marked stopReason: "error".

Steps to Reproduce

1. Use claude-opus-4-6 (streaming) in a coding-heavy session
2. Request code generation or file editing that may include control characters
3. Intermittently, a response will fail with the above error

Observed Behavior

• User sees raw JSON parse error as a chat message
• Response is lost (stopReason: "error")
• Session recovers on the next user message (no persistent corruption)
• Position in error varies (122, 131, 134, 156, etc.)

Expected Behavior

Either:

1. Retry the request automatically (the error is transient — next attempt usually succeeds)
2. Sanitize control characters from SSE data before JSON.parse() (defense-in-depth, since the upstream SDK should also fix this)
3. At minimum, show a user-friendly error like "Response interrupted, please try again" instead of the raw JSON parse error

Environment

• OpenClaw: 2026.2.9
• @anthropic-ai/sdk: 0.73.0
• Model: claude-opus-4-6 (streaming)
• Node.js: 22.x

Related

• Upstream SDK issue: U+2028 LINE SEPARATOR in MCP tool results causes JSON parsing failures. anthropics/anthropic-sdk-typescript#882 (U+2028 LINE SEPARATOR causes same class of JSON parse failure)
• OpenClaw Session corruption from invalid Unicode surrogates in browser tool output #7275 (similar: invalid Unicode surrogates in browser tool output)
• OpenClaw [Bug]: Calling opus4.6 resulted in an error, the JSON format could not be parsed and was truncated. #10507 (similar: Opus 4.6 JSON parse/truncation)
• OpenClaw Malformed tool call error leaked verbatim to user instead of being handled gracefully #7867 (similar: malformed tool call JSON leaked to user)