[Feature Request] Named, Replayable Agent Bootstraps (Procedural Checkpointing)

Summary
I’d like a way to define and reuse procedural Claude agents without relying on long-lived chat history or opaque session state.

Problem
For complex workflows (e.g. structured document extraction with validation), I carefully bootstrap a Claude instance with:
• A multi-step procedure
• Clear constraints and failure modes
• Verification steps
• A deliberately minimal, uncontaminated context

This setup works extremely well — but if the session resets, I lose the procedural state and must manually rebuild it through multiple turns.

I do not want conversational memory or indefinite context retention.
I want a way to reconstruct a clean agent state deterministically.

Proposed Solution
Introduce named agent bootstraps (or “agent profiles”) that allow users to:
• Define a reusable agent specification (role, constraints, workflow steps)
• Attach references (schemas, templates, examples)
• Re-instantiate this agent in a fresh session at any time
• Keep sessions short-lived and task-scoped
• Avoid carrying arbitrary conversational history forward

Environment Info

• Platform: darwin
• Terminal: WarpTerminal
• Version: 2.0.69
• Feedback ID: 098f0d00-2121-4b59-9934-2812161bbb59