[BUG] Task tool subagents permanently return 'Rate limit reached' with 0 tokens consumed

Resolved 💬 7 comments Opened Feb 20, 2026 by MalekAG Closed Mar 23, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

The Task tool (subagent spawning) permanently returns API Error: Rate limit reached with total_tokens: 0 and tool_uses: 0. This persists across multiple days despite having available usage/credits on the Anthropic account.

Key observations:

  • Affects all subagent types (Explore, general-purpose, etc.)
  • Affects all models (Opus, Sonnet, Haiku)
  • total_tokens: 0 indicates the request is rejected before any tokens are processed
  • The main session (parent agent) works fine — only Task tool subagents are affected
  • Has persisted for days, not minutes — this is not a transient rate limit
  • Account has remaining usage/credits

Example output:

API Error: Rate limit reached
<usage>total_tokens: 0
tool_uses: 0
duration_ms: 296</usage>

This appears related to #25778 (misleading rate limit when safety filter triggers) and #25531 (false rate limit at context boundary), but differs in that:

  • It only affects subagent spawning, not the main session
  • It persists across completely new sessions and days of waiting
  • /clear does not resolve it

What Should Happen?

Task tool subagents should spawn normally when the account has available usage. If there's an actual error (safety filter, context issue, etc.), the error message should reflect the real cause, not a generic "rate limit reached."

Steps to Reproduce

  1. Use Claude Code with Task tool to spawn subagents
  2. At some point, all subagent calls start returning "Rate limit reached" with 0 tokens
  3. Wait days — error persists
  4. Start new sessions — error persists
  5. Try different models (Opus, Sonnet, Haiku) — all fail identically

Claude Model

Opus 4.6 (main session), tested subagents on Sonnet and Haiku as well

Is this a regression?

Yes — subagents worked previously

Claude Code Version

2.1.49 (Claude Code)

Platform

Windows

Operating System

Windows 10 Pro 10.0.19045

Terminal/Shell

Bash (Git Bash)

Additional Information

The 0 total_tokens strongly suggests this is not a real rate limit — a genuine rate limit would still count input tokens before rejecting. This looks like the API is returning a rate limit error code for a different underlying issue (possibly related to the safety filter masking described in #25778).

View original on GitHub ↗

This issue has 7 comments on GitHub. Read the full discussion on GitHub ↗