API 500 errors fail silently, stalling sessions with no recovery

Problem

During a normal Claude Code session (building a multi-file Python project with background agents), I'm hitting recurring API 500 Internal Server Errors. The core problem isn't the 500s themselves — those happen — it's that Claude Code has no error handling for them. When a 500 occurs:

• The agent/tool call just silently stops
• Claude doesn't know it failed
• The session stalls with no indication of what happened
• The user has to notice the stall, manually check, and nudge Claude to retry
• Work that was in progress is lost with no recovery

This makes Claude Code very difficult to use for any non-trivial session. Background agents are especially affected since there's no feedback loop.

Request IDs

These all occurred within a single ~30 minute session on February 25, 2026:

• req_011CYVAjspabZ5YKNPgKUTnf
• req_011CYVAfLw6t9VeQqNroHPw9
• req_011CYVAeCxrGK9dCAz16tpaM
• req_011CYVCFPVCzRQSbv5qVf3as

Expected Behavior

• Claude Code should catch 500 responses and retry automatically (with backoff)
• If retries fail, surface the error clearly to both the user and the Claude context
• Claude should be told "your last tool call failed with a 500, retry or adjust"
• Background agents should report failures back to the parent session
• At minimum: don't fail silently. Show the user something.

Environment

• Claude Code CLI
• Model: claude-opus-4-6
• macOS (Darwin 25.3.0)
• Using background agents (Task tool) extensively

Impact

In a session building a multi-file project, 4 separate 500 errors caused the session to stall repeatedly. Each time required manual intervention to notice, diagnose, and recover. For a tool that's designed for autonomous multi-step work, silent failures are a critical gap.