Sub-Agent Delegation with Context Isolation

PROJECT CONTEXT:
My agent's context fills with exploration noise (search results, file dumps) that crowds out the actual task. Analyses of Claude Code show its answer: spawn sub-agents that explore in their own context and report back only conclusions.

GOAL:
Implement a dispatch tool: the main agent delegates a bounded task to a sub-agent with a fresh context, and receives only the final distilled report.

REQUIREMENTS:
- A dispatch_agent tool: task description in, final report out
- Sub-agent runs its own loop with its own message history
- Sub-agent gets a restricted tool set (read-only by default)
- Only the sub-agent's final report enters the parent's context
- Depth limit (no sub-sub-agents in v1) and turn budget per sub-agent

CONSTRAINTS:
- Sub-agent transcripts are logged but never injected into the parent
- The parent must phrase tasks with explicit deliverables, not vague asks

DEFINITION OF DONE:
- A codebase question is answered via sub-agent with under 500 tokens entering parent context
- Sub-agent cannot call write tools when dispatched read-only
- Sub-agent hitting its turn budget returns a partial report, not an error
- Parent transcript shows only task + report, never raw exploration

COMMON FAILURES TO AVOID:
- Returning the sub-agent's whole transcript, defeating the purpose of isolation
- Vague task prompts producing vague reports the parent cannot use
- Unlimited recursion: sub-agents spawning sub-agents until the budget burns
- Giving sub-agents full write access for what should be a research errand