Codex Gives Skill Subagents Clean Context

This is a small runtime change with outsized impact: skills are only as reliable as the context they enter.

• –Fresh contexts reduce instruction drift and accidental carryover from the orchestrator’s conversation.
• –Code-review agents can focus on repository evidence instead of inheriting prior assumptions.
• –The trade-off is repeated setup and potentially higher token usage for each subagent.
• –Codex documentation already emphasizes separate subagent threads, making consistent context isolation an important implementation detail for skill authors.