Feature request: first-class agent review-gate primitive for multi-agent workflows

I've been running a manual project-manager/implementer pipeline with Claude Code that exposes a missing primitive worth building into the SDK or product.

The workflow:

1. A Claude Opus agent (acting as project manager) writes a spec — a structured markdown file with verification greps, exact code targets, and a Do-Not-Change fence.
2. A human reviewer approves or rejects with feedback.
3. A Claude Sonnet agent (acting as implementer) executes only on approval.

This maps naturally onto the existing multi-agent SDK — except step 2 requires me to act as the router. There's no native primitive that says: pause execution, send output to a named reviewer agent, block until it approves, then unblock the next agent.

What would make this a product:

• A review_gate tool or step type in the Agent SDK: routes output to a specified agent (or human), carries approval/rejection + feedback back to the pipeline, and only proceeds on approval.
• A reject_with_feedback return path that feeds corrections back to the originating agent without restarting the whole pipeline from scratch.
• Optional: named agent roles with enforced tool restrictions per role — e.g. the reviewer cannot edit files, the implementer cannot write specs.

The manual version of this works well — we've shipped ~93 iterations this way. The friction is the human-in-the-loop handoff that could be automated. Anthropic has all the pieces; they just aren't wired together as a first-class workflow primitive.