Graceful degradation on internet disconnect during orchestrator execution

Problem

When internet connectivity drops during orchestrator execution, the system degrades ungracefully:

• Gemini cache TTL expires and cannot refresh → all subsequent calls uncached (10x cost)
• fetch failed errors cascade through inspector cache creation
• Subagent announce queue times out on unreachable API
• Orchestrator session ends, restart loop fires, but model stalls due to degraded API

Evidence

E2E test on 2026-03-10: Internet crashed at ~06:21 UTC. Before crash: 48 cache hits, wave 1 complete, wave 2 in progress. After crash: cache lost, session 2 model stalled (3 follow-up exhaustions), likely due to intermittent Gemini API.

Expected Behavior

• Cache should survive brief outages (retry refresh, extend TTL)
• LLM call failures should be retried with backoff (not silent text-only fallback)
• Orchestrator should checkpoint and pause gracefully on sustained connectivity loss
• On reconnect, resume from checkpoint without losing cache

References

• Test report: docs/plans/2026-03-10-session-persistence-e2e-report.md
• Gateway log evidence: fetch failed at 06:21:03, 06:21:38

🤖 Generated with Claude Code