[BUG] Bug: Rate limit errors on Pro Plan at 26% usage — rate limit stuck per-session, contradicts documentation

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?

I am a paying Pro Plan subscriber being blocked with "API Error: Rate limit reached" while my usage dashboard shows only 26% daily usage (monthly is even lower).

Critical finding: while the session was blocked, I opened a NEW Claude Code session on the same machine, same VSCode instance, same Pro account — and it worked perfectly. This proves the rate limit is stuck per-session, not per-account, which directly contradicts Anthropic's own documentation stating limits are applied "at the organization level."

Root cause from session logs: Claude Code launched a burst of ~10 parallel tool calls (glob + read operations) in 30 seconds. 11 seconds after the last successful response, the rate limit triggered. The agent's own parallel execution strategy caused the rate limit — not the user.

Additionally, the VSCode extension UI displayed a "stopped responding" message during parallel agent operations, prompting the user to cancel running tasks. This misleading UI message compounds the problem.

What Should Happen?

1. A user at 26% daily quota should NOT be blocked by rate limits
2. If a short-window throttle (RPM) is hit, the error message should say WHICH limit was reached and WHEN it resets — not a generic "Rate limit reached"
3. Rate limits should apply at account level, not get stuck on a single session
4. Claude Code should manage its own parallel tool calls to avoid triggering RPM limits
5. The VSCode UI should not display "stopped responding" when parallel operations are still running

Error Messages/Logs

{
  "type": "assistant",
  "timestamp": "2026-03-31T16:19:39.213Z",
  "message": {
    "model": "<synthetic>",
    "content": [{"type": "text", "text": "API Error: Rate limit reached"}],
    "usage": {
      "input_tokens": 0,
      "output_tokens": 0,
      "cache_creation_input_tokens": 0,
      "cache_read_input_tokens": 0
    }
  },
  "error": "rate_limit",
  "isApiErrorMessage": true,
  "version": "2.1.84"
}

Note: input_tokens: 0 — the request was rejected before processing any tokens.

Steps to Reproduce

1. Subscribe to Pro Plan
2. Use Claude Code via VSCode extension for a multi-day session (mine ran 7 days)
3. Ask Claude to perform parallel file operations (e.g., reading multiple PDFs or globbing large directories)
4. Claude launches a burst of ~10 parallel tool calls in ~30 seconds
5. Rate limit triggers 11 seconds after the last successful response
6. ALL subsequent messages return "API Error: Rate limit reached"
7. Dashboard shows ~26% daily usage
8. Open a NEW Claude Code session on the same machine, same account — it works immediately (proving the limit is per-session, not per-account)

Session ID: 63b58d02-fd89-44cb-a42b-f88c120f2e38
Token consumption: ~4.4M total (1,870 input, 13,960 output, 371,320 cache creation, 4,010,843 cache read)

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.84

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

Session token analysis:
| Metric | Value |
|---------------------|-------------|
| Input tokens | 1,870 |
| Output tokens | 13,960 |
| Cache creation | 371,320 |
| Cache read | 4,010,843 |
| Total | ~4.4M |
| Successful messages | 89 |
| Rate limit errors | 5 |

Timeline:
| Event | Timestamp (UTC) |
|--------------------------|-------------------------|
| Session start | 2026-03-24 17:04 |
| First rate limit error | 2026-03-24 18:50 |
| Last successful response | 2026-03-31 16:12 |
| Final block (4 errors) | 2026-03-31 16:12—16:19 |

The full analysis is documented in the session JSONL file. I can provide it if needed.

This is a contractual concern: I paid for Pro Plan capacity and I am being blocked at 26% of that capacity. The rate limiter contradicts Anthropic's own documentation. I expect transparency, accurate error messages, and a fix.