The tiny status label that rereads a Claude Code agent's entire history

Claude Code can run background agents: helpers that work on a job on their own while the main conversation goes on. Each one shows a short status label, three to five words on what it is doing. On October 26, a company called Flamerix AI published a write-up saying that label is not free. About every 30 seconds, for each working helper, Claude Code sends the helper's whole transcript (everything it has read and written so far) back to the model and asks for those few words.

Flamerix says that in one measured week, these label calls were 16.4% of its estimated Claude Code spend and 19.7% of its requests. On October 29, a member of the Claude Code team answered on Hacker News. They confirmed how it works, but said it only happens for helpers that are actively working, not idle ones. Across users who share data with Anthropic, they put it at about 2% of cost last week, and about 6% in the terminal. They said a fix for the most expensive case, labels that miss the prompt cache, shipped in version 2.1.311. A cache lets the model reuse text it has already read, which costs far less than reading it fresh. They added that a cheaper mode, which sends short notes to a smaller model instead of the whole transcript, is in the works.

This touches Jon's setup directly. Claude Code is his main tool, and since October 28 his fleet hands most jobs to background helpers. A re-measure the fleet ran that day found a helper job cost the lead session about 0.36 million weighted tokens, against 0.75 million for a linked session and 1.67 million for a Ringer run. Tokens are the small chunks of text a model reads and writes, and they are what fills a plan's usage limit, the cap on how much work a subscription allows in a stretch of time. Each working helper also pays for its label. One fleet machine already runs Claude Code 2.1.312, so it has the cache fix.

How to use it

Update to Claude Code 2.1.311 or later. Run claude --version to check, and claude update to move up. The 2.1.311 changelog also fixed /usage so it now shows large prompt-cache misses under "What's contributing to your limits usage?"

To see the label's real share, turn on Claude Code's OpenTelemetry export, its built-in usage logging. Group the api_request events by query_source and look for agent_summary. Their cost_usd is an estimate, not a bill; on a subscription, what is lost is room under the usage limit.

There is no official off switch. Flamerix's script blocks the label timer, but it is unofficial, tested on Linux only, and an update can quietly undo it. A GitHub commenter warns that it pulls the newest copy each time and runs inside every Claude Code process. The same thread says setting CLAUDE_CODE_FORK_SUBAGENT=0 stops the labels too, but the docs only say it controls fork mode, which it turns off.

What nobody knows yet

Which number fits a given workload: Flamerix ran mostly background helpers, and Anthropic's 2% is an average over everyone. The fleet's own re-measure counted the lead session's side of each job and did not separate out label calls, so the fleet has not measured this cost yet. The changelog for 2.1.311 does not name the label fix.

What to watch

The cheaper smaller-model mode, and whether it ships with a setting to turn the labels off. The GitHub request for a supported switch has been open since August 15 and has no answer from Anthropic yet.