1M-context session becomes unrecoverable: ECONNRESET on large uncached requests, and /compact fails with the same error

Summary

Sessions with very large 1M-context conversations (~520k tokens) repeatedly fail with API Error: Unable to connect to API (ECONNRESET) when the prompt cache has gone cold, and — critically — /compact fails with the same error, because compaction itself requires sending the full context. This leaves the session unrecoverable: the user's only option is abandoning it and starting a new session, losing all conversational state.

Environment

• Claude Code v2.1.197, macOS (Darwin 25.2.0)
• Model: claude-fable-5[1m] (1M context beta)
• No VPN, no proxy, no ANTHROPIC_BASE_URL override
• status.claude.com showed all systems operational during the failures

Symptoms

• Affected session context: ~520k tokens; last successful request showed cache_creation_input_tokens: 496847 (i.e., nearly the entire context re-ingested uncached after the ~5 min cache TTL lapsed).
• Requests hang for many minutes ("Cogitated for 8m 7s") then die with ECONNRESET or "socket connection was closed unexpectedly"; occasional overloaded_error (529) and 500s in the same windows.
• Errors came in bursts across multiple days (2026-07-04 16:08–16:27, 20:50, 21:32–22:03 UTC; 2026-07-05 18:14+ UTC).
• Small sessions (~45k tokens, mostly cache hits) in the same environment at the same time worked flawlessly. Direct connectivity checks to api.anthropic.com during a failure window were clean (5/5 connects, ~27ms).
• Attempting to recover with /compact failed: Error during compaction: API Error: Unable to connect to API (ECONNRESET).

Why this is a product bug and not just transient API weather

1. Auto-retry re-pays the full ~500k-token uncached ingestion on every attempt, keeping the session in the failure-prone state for many minutes per retry cycle.
2. /compact — the documented remedy for oversized context — requires the exact same giant request that is failing. When a session most needs compaction, compaction is least likely to succeed.

Suggested improvements

• Make compaction resilient for huge contexts: chunked/incremental summarization, or compaction that can resume across failed attempts instead of resending the entire context each time.
• Surface a clearer error/hint when repeated ECONNRESETs correlate with very large uncached prompts (e.g., suggest compacting before the context grows past the point where compaction itself can fail).
• Consider keeping the prompt cache warm (or warning on cache expiry) for sessions above some context size, since a cold cache is what turns each request into a multi-minute vulnerable window.

Impact

User lost an active large working session and had to reconstruct state in a fresh session by hand. Retries did eventually succeed for normal messages, but each cycle cost many minutes, and the recovery path (/compact) was unusable.