Background agents hang silently + no incremental output for large file generation

Problem

When using Claude Code with Max subscription, background agents (run_in_background: true) hang silently without producing output, errors, or timeouts. The user has no visibility into agent progress.

Steps to Reproduce

1. Launch a subagent with run_in_background: true for a large task (e.g., generating a 1500+ line HTML file with detailed SVG illustrations)
2. Agent never completes — no output file, no error, no timeout
3. Checking the output file shows empty content
4. No way to kill or restart the stuck agent

Additional Issues

• No incremental output: The Write tool requires full content generation before writing. For large files (1000+ lines), the user sees no activity for 1-2 minutes, making it appear Claude is frozen.
• User interrupts reset everything: If the user interrupts during long generation, all progress is lost. No checkpoint mechanism exists.
• No agent health monitoring: Cannot check if a background agent is working, stuck, or failed.

Expected Behavior

• Background agents should have a timeout (e.g., 5 minutes) and surface errors
• Agent progress should be visible to the parent context
• Large file writes should support incremental/streaming output
• Stuck agents should be killable

User Feedback

I pay for the Max subscription (300€/month) and expect reliable, professional-grade performance. This session resulted in 20+ minutes of waiting with no deliverable produced due to silent agent failures. The experience felt like being strung along with repeated "sorry, let me try again" responses while nothing actually worked. This is unacceptable for a premium product. I expect competence and reliability, not excuses.

Environment

• Claude Code with Anthropic Max subscription
• Model: claude-opus-4-6
• Platform: Windows 11
• Multiple agent types tested (visual-architect, jamstack-ninja, nodejs-ninja, tailwind-artist)

Impact

Complete failure to deliver requested output. User trust severely damaged. The tooling around background agents is not production-ready and should either be fixed or disabled until reliable.