API rate limits when Skill/slash command is invoked early in conversation with Opus model

Description

When using Claude Opus 4.6 as the default model, invoking a skill or slash command early in a conversation consistently triggers API rate limit errors. The same skill works fine when:

1. The skill/slash command is invoked later in the conversation (after some prior context)
2. The skill is triggered indirectly via natural language (causing the Skill tool to be called as part of a larger response)

Steps to Reproduce

1. Set default model to Claude Opus 4.6
2. Start a new conversation
3. Immediately invoke a skill or slash command (e.g. /commit, /pr, or a custom skill)
4. Observe API rate limit error

Expected Behaviour

The skill should execute without rate limit errors, regardless of when in the conversation it is invoked.

Actual Behaviour

Rate limit error is returned when the skill is the first (or very early) action in a conversation.

Environment

• Platform: macOS (Darwin 25.3.0)
• Model: Claude Opus 4.6 (1M context)
• CLI: Claude Code

Notes

This may be related to how skill expansion + Opus token usage hits the rate limit window when there is minimal prior conversation context. When the skill is invoked later or indirectly, the token usage may be spread differently across the rate limit window.