Task agents should inherit parent conversation context

Problem

When using Task agents (subagents) for parallel work, the agents start fresh without the parent conversation's context. This means they miss:

• User preferences and design decisions discussed earlier in the conversation
• Nuances about what's already been tried and rejected
• Specific feedback the user gave about previous iterations
• Understanding of the codebase built up through the conversation

This leads to agents producing work that doesn't align with what the user wants, requiring re-work in the main conversation anyway.

Suggestion

Task agents should be forks of the current context — they should inherit the full parent conversation history (or a summarized version of it) so they have the same understanding as the main conversation. This would make them much more effective for parallel work since they'd understand:

• What the user has already said they like/dislike
• Design decisions and constraints discussed
• Previous iterations and why they were rejected
• The overall direction of the work

Current Workaround

Avoiding agents entirely and doing all work directly in the main conversation, which loses the parallelism benefit.

---
Reported via Claude Code CLI