feat: vision/image support via stream-json + CLI isolation

[rjabalosiii]

Summary

Adds multimodal (text + image) support for OpenAI-compatible clients that send base64 screenshots, such as Browser Use. Also fixes several compatibility issues discovered during real-world testing.

What changed

• Vision/image support: Auto-detects image_url content parts in OpenAI requests and switches to --input-format stream-json mode, piping NDJSON with base64 image blocks via stdin. Text-only requests still use the fast CLI argument path (zero overhead when no images).
• Multimodal content arrays: Handles OpenAI's content field as both string and Array<{type, text?, image_url?}> — required for any client sending screenshots or structured content.
• Code fence stripping: Claude wraps JSON responses in markdown code fences (\``json...). Added stripCodeFences()` to extract clean JSON for clients expecting raw JSON (e.g., Browser Use's structured output).
• CLI isolation: Prevents Claude Code's agentic system prompt, tools, skills, and settings files from leaking into proxy responses. Flags: --tools "", --disable-slash-commands, --setting-sources "", custom --system-prompt, cwd: /tmp.
• Body limit bump: 10mb → 50mb to accommodate base64-encoded screenshots.

How it works

Client sends image_url (base64 data URI)
  → Proxy detects hasImages=true
  → Converts to Claude CLI stream-json format:
    {"type":"user","message":{"role":"user","content":[
      {"type":"text","text":"..."},
      {"type":"image","source":{"type":"base64","media_type":"image/png","data":"..."}}
    ]}}
  → Pipes via stdin with --input-format stream-json
  → Claude CLI processes multimodal input
  → Response streamed back as SSE chunks

Backward compatibility

• No breaking changes — text-only requests follow the same code path as before
• Stream-json mode only activates when images are detected

Test plan

• E2E tested with Browser Use v0.11.9 (vision-enabled browser automation)
• Stress tested with 3 task types: data extraction, multi-page navigation, structured output
• Verified text-only requests still work via CLI argument path
• Verified base64 PNG screenshots pass through to Claude and get vision responses

🤖 Generated with Claude Code