[Bug] 72% of API requests are empty polling requests counting toward usage quota

Bug Description
Bug: 72% of API requests are empty polling that appears to count toward usage quota

SETUP:
I built a local HTTP proxy (Node.js) to intercept and log ALL requests between Claude Code and Anthropic's API to investigate why my usage is draining abnormally fast.

ENVIRONMENT:

• Windows 11, PowerShell, VS Code
• Claude Code v2.1.91 (also tested on stable v2.1.84 - same results)
• Max plan
• Model: Opus 4.6 (1M context)
• Effort: low
• All plugins and agents DISABLED for clean test

TEST: Asked Opus to implement an already-generated plan (a routine task that normally takes ~10 min and ~1% usage).

PROXY RESULTS (34-minute session):

• Total API requests to /v1/messages: 94
• Requests with actual token usage (input or output > 0): 27 (29%)
• EMPTY requests (0 input, 0 output): 68 (72%)
• Additional count_tokens requests: 7
• Total cache read: 2.18M tokens
• Total cache write: 33.5k tokens
• Total input tokens: 2.4k
• Total output tokens: 7.7k
• Thinking tokens: 0
• Usage consumed: ~7% of 5h session

COMPARISON WITH PLUGINS ENABLED (same type of task, 18 min):

• Total requests: 111
• Empty requests: 92 (83%)
• Cache read: 1.99M
• Usage consumed: 3-4% (on Sonnet)
• Tool count per request: 21 (from plugin agents)
• Message count grew from 5 to 225 in a single task

KEY FINDINGS:

1. 68-92 empty polling requests per task that appear to count toward quota
2. These empty requests continue even when user is idle (not typing anything)
3. Last 5 requests before I stopped were ALL empty - Claude Code keeps polling after task completion
4. Only 27 of 94 requests did actual work - the rest are pure overhead
5. Disabling all plugins reduced requests by ~40% but empty polling persists
6. Cache read accumulates to 2.18M for a simple implementation task
7. Yesterday (April 2) identical workload with identical setup did NOT have this issue
8. Tested on v2.1.84 (stable) - same behavior, confirming this is server-side

IMPACT:

• Tasks that used to cost ~1% now cost 7%
• Session went from 19% to 43% usage with just a few tasks
• At this rate, Max plan provides ~2 hours of work instead of 5

I have the full JSONL proxy logs with timestamps, token counts, model, messageCount, toolCount for every request. Happy to share if helpful for investigation.

Related issues: #43183, #43185, #43198, #41788

Environment Info

• Platform: win32
• Terminal: windows-terminal
• Version: 2.1.91
• Feedback ID: 0acd3619-dba5-4d17-b88b-07bb597c51ae