Escape Unicode separators in JSON-RPC lines

[georgecollier-nqu]

Current failure

BB uses one JSON object per line between the host daemon and each provider bridge.

Injected instructions can contain U+2028 or U+2029. These are valid Unicode line separator characters.

1. JSON.stringify leaves each character literal inside the JSON string.
2. Node readline treats each character as the end of a line.
3. A thread/start request becomes two incomplete lines before the bridge parses it.
4. The bridge catches each JSON.parse failure and returns without a response.
5. The host waits until thread/start times out. Claude never starts.
6. The user can remain on Preparing workspace without a useful error.

This failure occurs before BB starts Claude. It is a BB message framing bug, not a Claude Code failure.

Why BB must fix this

The input text is valid and must remain valid. It can come from these sources:

• AGENTS.md
• skills
• system instructions
• prompts
• tool schemas
• provider messages

Sanitizing each source would be fragile and could change user text. The JSON-RPC framing layer owns this requirement.

Escaping these characters keeps each message on one physical line. JSON.parse restores the original Unicode text at the receiver.

Fix

• Escape U+2028 and U+2029 before BB writes line-delimited JSON-RPC.
• Apply the fix to runtime-to-bridge messages.
• Apply the fix to bridge-to-runtime messages.
• Add regression tests that pass both characters through Node readline.

Validation

• pnpm exec vitest run src/runtime-json-rpc.test.ts --config vitest.config.ts
• pnpm exec turbo run typecheck --filter=@bb/agent-runtime

AGENT GENERATED: by GPT-5.6 Sol

[georgecollier-nqu]

reviewed manually too