Claude Code regularly sends each background agent’s transcript to its model(gist.github.com)
469 points by NikZaal 10 hours ago | 187 comments
tl;dr: Claude Code silently burns ~15-20% of user tokens (mostly on Opus) by having each background agent summarize its own transcript every ~30 seconds into a 3-5 word status line that never leaves the agents panel. There's no official setting to disable it, so the author ships an ~80-line Bun preload hook that wraps setTimeout/setInterval and suppresses the summarizer timer by matching its callback source text. Tested on Linux across Claude Code 2.1.257–2.1.312; it may silently fail after updates, so re-verification via OpenTelemetry is recommended.
HN Discussion:
  • ↑Confirms the token-waste problem with personal usage estimates supporting the article's claim
  • ↑Appreciates the cleverness of the setTimeout-patching hack as true hacker spirit
  • ↑Broader criticism that Claude Code is becoming bloated and inference providers shouldn't own harnesses
  • ↓Questions the premise, suggesting transcript summaries would mostly be cheap cache hits
  • •Adds related information about other Claude Code issues like email in headers or cache invalidation