Claude Code re-bills a finished sub-agent's whole context on follow-up, wasting quota
Schadz · reddit · 2026-09-16
A Reddit user diagnosed a Claude Code quota anomaly: when you send a follow-up to a finished sub-agent, Claude Code rebuilds its conversation in a way the prompt cache doesn't recognize (messageschanged), re-writing nearly the agent's entire context at expensive cache-write rates. In tests, two resumed agents wrote 243K and 399K tokens back to cache, versus only 2K-4K per message while agents were still running.
Key findings:
- The miss happens inside the 1-hour cache window, so it's not expiry; the API explicitly reports messageschanged
- Comparing request sizes shows the changed content is the agent's thinking: 312,205 tokens dropped to 259,599, matching exactly 53,084 thinking tokens (within 0.07%)
- Sub-agents with thinking always miss on resume; agents with no thinking got a full cache hit
- The shape resembles Anthropic's April 23 postmortem about a client-side change that cleared thinking on idle sessions
The author has filed the issue with Anthropic and urges affected users to check their own usage.
More from coding & agent
- Microsoft's MemSteer turns unlabeled user histories into agent memory supervision — MicrosoftResearch · 2026-11-02
- She fired herself from her own Instagram: AI agent runs the account, views up 82.3% — heyshrutimishra · 2026-11-02
- Sourcegraph: AI writes code faster than teams can review it, driving more outages — AI Engineer · 2026-11-02
- Tsinghua & Microsoft's HarnessMM cuts agent token use by up to 87.93% while raising success rate — rohanpaul_ai · 2026-11-02
- GitHub launches Agent HQ: Anthropic, OpenAI, Google coding agents in Copilot — emmanuelvivier · 2026-11-02
- DeepSeek open-sources Super-Agent, a coding agent that reads team chats to minimize human input — emmanuelvivier · 2026-11-02