Sonnet dispatched via Agent tool fails with "1M context required" error

Description

Calling the Agent tool with model: "sonnet" from a Claude Code session (parent: Opus 4.7) fails immediately with:

API Error: Usage credits required for 1M context · run /usage-credits to turn them on, or /model to switch to standard context

This happens regardless of prompt size — even a 5-token "reply ok" prompt fails at 0 tokens consumed.

Reproduction

1. In a Claude Code session with Opus 4.7 as parent model
2. Invoke the Agent tool with subagent_type: "general-purpose", model: "sonnet", prompt: "Reply ok"
3. Immediate error before any tokens are consumed

Expected behavior

Standard-context Sonnet dispatch (200k, or 600k for users in that A/B). This pattern worked previously without requiring 1M context credits.

Actual behavior

Sonnet dispatch is gated behind 1M context credits, which must be explicitly enabled and incur additional cost.

Diagnostic

• Haiku dispatch: works normally (~38k tokens for a small test, no credit error)
• Sonnet dispatch: blocked immediately at 0 tokens with the 1M error
• No local settings (settings.json, settings.local.json, or any CLAUDE.md on the path) reference context size or 1M

Environment

• OS: Windows 11, x64
• Claude Code version: 2.1.144
• Parent model: Opus 4.7
• Flags in use: --remote-control, --dangerously-skip-permissions
• Interactive mode only (no openclaw / non-interactive automation)

Impact

Breaks the documented "Opus orchestrates, Sonnet codes" pattern from the Agent tool description. Either:

1. This is intended policy and should be documented (with updated cost guidance for the orchestration pattern), or
2. This is an accidental flag and Sonnet dispatch should not require 1M context by default.