T3 Review Quorum

A Code Quorum review whose reviewers run as T3 delegated tasks on two model families. Read and follow ~/.agents/skills/code-quorum/SKILL.md for the whole workflow: its read-only boundary, scope, reviewer selection, finding contract, verification, and report template all apply unchanged. This skill replaces only the worker mechanism and the reviewer models.

Confirm agents

This step replaces Code Quorum's Confirm agents step.

The worker mechanism is T3's delegate_task. Each delegated task is a blind agent: it starts with no conversation history and sees only its task packet.

The reviewer models are Astra 6 and Fable 5.1. When the user names other models, use those instead. Call orchestrator_capabilities and resolve each model to an exact providerInstanceId and model id from the live catalog. Prefer the provider's native driver (codex for OpenAI models, claudeAgent for Claude models) over a proxy such as Pi, and pick only an instance with canRunChildTask: true. When a model is unavailable, tell the user which one is missing and ask how to proceed.

Complete this step when every reviewer model resolves to a providerInstanceId and model id that can run a child task.

Run the quorum

This step replaces how Code Quorum's Run the quorum step creates agents; its task packet, waiting rules, and failure rules still apply.

Run every selected rubric once on each reviewer model, so each rubric gets a reviewer from both families. For each pair, call delegate_task with: