A hardcoded instruction in Claude Code 2.1.219 and 2.1.220 blocks Opus 5 from spawning sub-agents

According to a report on GitHub, a system prompt section labeled heron_brook is hardcoded to restrict Opus 5. Two lines explicitly forbid the model from calling AgentTool or using workflows and deep-research unless the user explicitly requests them.

This affects skills that rely on sub-agent delegation. The model reportedly performs these tasks within the main session or skips them entirely without warning, making the output appear as a standard execution. In one instance, a self-audit failed to trigger an independent agent, with the model simply noting the issue and moving on. You can check your local environment for this restriction by searching for the string heron brook in ~/.claude.json.