Interactive session turns hang intermittently on all models (Opus 4.8, Sonnet 4.6) — headless calls fast

Summary

In an interactive Claude Code session, inference turns frequently hang (stall for 30s–several minutes, or indefinitely until interrupted) on all models tested, including claude-opus-4-8 and claude-sonnet-4-6. The same API path works fine from a headless claude -p call.

Environment

• Claude Code version: 2.1.156
• Node.js: v25.8.1
• OS: macOS Darwin 25.3.0, arm64 (Apple Silicon)
• Model(s) tested: claude-opus-4-8, claude-sonnet-4-6
• Context size: ~77k tokens (well under 1M limit)

Behaviour

• Interactive turns hang intermittently — sometimes for tens of seconds, sometimes until the user interrupts
• Occurs on both claude-opus-4-8 (newly rolled out) and claude-sonnet-4-6, ruling out a single-model issue
• Occurs in fresh sessions as well as long ones, ruling out context size
• Starting a new session does not resolve it
• The hang looks like waiting for first token (no output at all), not a slow stream

What works fine

Headless claude -p "reply OK" call in the same terminal:

• Opus 4.8: 200 in 1542ms, zero retries
• No overload, no 429/529 in debug logs
• Rate limits barely touched: 49/50 requests, 486k/500k input tokens remaining

Suspected cause

The interactive session sends a much larger set of beta headers than headless:

claude-code-20250219, context-1m-2025-08-07, interleaved-thinking-2025-05-14,
thinking-token-count-2026-05-13, context-management-2025-06-27,
prompt-caching-scope-2026-01-05, mid-conversation-system-2026-04-07,
advisor-tool-2026-03-01, advanced-tool-use-2025-11-24, effort-2025-11-24,
cache-diagnosis-2026-04-07

Headless uses a much smaller set. One of these beta flags may be triggering a slow/stalled server-side code path intermittently.

Debug output (working headless call for reference)

[log_ad9e50, request-id: "req_011CbVeji1QK5bZFDKXM3FsK"] post https://api.anthropic.com/v1/messages?beta=true succeeded with status 200 in 1542ms
anthropic-ratelimit-requests-remaining: 49
anthropic-ratelimit-input-tokens-remaining: 486000

Reproduction

1. Open an interactive Claude Code session (claude in terminal)
2. Ask any question that triggers an inference turn
3. Observe: ~50% of turns hang with no output until interrupted
4. Run claude -p "reply OK" in a separate terminal — returns in ~1.5s