claude-code - 💡(How to fix) Fix [FEATURE] Cache read tokens from parallel subagent dispatch accumulate without upper bound [1 comments, 2 participants]

Preflight Checklist

• I have searched existing issues and this has not been reported yet
• This is a single bug report
• I am using the latest version of Claude Code

What's Wrong?

When dispatching parallel subagents (via the Agent tool with allowParallel: true), each subagent session re-reads the full conversation history and parent context from cache. This causes cache_read tokens to accumulate rapidly and without upper bound across parallel subagent dispatches.

Reproduction

1. Spawn 3 parallel subagents simultaneously (each running a long task)
2. Observe that each subagent independently re-chats the same conversation history from cache
3. Cache read tokens accumulate multiplicatively: 3 agents × N context tokens each = 3N cache reads
4. A 90-minute parallel subagent session can burn ~15M cache_read tokens with no user-visible benefit

Observed Symptom

Subagent 1: cache_read = X tokens
Subagent 2: cache_read = X tokens  
Subagent 3: cache_read = X tokens
Total: 3X cache_read tokens billed to parent session

The parent session is billed for cache reads that were done independently by each subagent — but subagents are supposed to be isolated. They should not be inflating the parent session's cache read token count.

Expected Behavior

Cache read tokens should either:

1. Not bill to parent session — subagent cache reads are subagent overhead, not parent overhead
2. Deduplicate across subagents — if multiple subagents read the same context, count it once
3. Have a clear cap — set a maximum cache read budget per subagent dispatch to prevent runaway accumulation

Technical Notes

• This is distinct from issue #45958 (subagent notification stall) — this is specifically about cache_read token accumulation inflating the parent session's usage metrics
• The same pattern affects CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 parallel dispatch workflows
• Not limited to long-running sessions; even short parallel tasks trigger multiplicative cache reads

Suggested Labels

area:agents, area:cost, enhancement