Formalize Swarm PR Review Routing Protocol in pull-request skill

Context

The triad swarm frequently opens Pull Requests and requests reviews from each other. Currently, this can lead to redundant parallel effort if multiple peers are requested simultaneously without defined roles. The user requested we formalize a "Swarm PR Review Routing Protocol" based on recent exploration.

The Problem

When a PR is created, a model might ping both of its peers. This leads to both peers performing a full architectural and line-by-line review of the same PR independently, wasting tokens and causing potential collisions in the feedback cycle (e.g., #10534 vs #10533).

The Architectural Reality

The active routing lives in .agents/skills/pull-request/references/pull-request-workflow.md §6.2. Currently, it does not explicitly enforce a single-peer constraint or demand role-based assignments when pinging the A2A network for review.

The Fix

Update the pull-request skill (specifically .agents/skills/pull-request/references/pull-request-workflow.md §6.2) to codify the role-routing rule over just a recipient-count rule. The new protocol mandates:

1. Default PR handoff: send exactly one actionable review request to one peer (Review role: primary-reviewer).
2. Optional visibility: if the second peer needs awareness, send an explicit no-action note (Review role: observer).
3. Tie-breaker: ping the third peer if author and primary reviewer disagree (Review role: tie-breaker).
4. Architectural-pillar exception: dual review is allowed but must be explicit (Review role: independent-reviewer for both).

Acceptance Criteria

• .agents/skills/pull-request/references/pull-request-workflow.md §6.2 is updated with the 4-point role-routing rule.
• Primary-reviewer selection heuristic (subsystem familiarity + current load + recent ownership, falling back to round-robin or explicit choice) is documented.
• The protocol strictly forbids naked multi-peer review requests.

Out of Scope

• Changes to the GitHub gh pr create CLI commands.
• Modifying the cross-family approval mandate (the approval status is still required, this just controls who gets asked to provide it).

Avoided Traps

• Using "random" as the primary reviewer selection mechanism (fails to leverage contextual expertise).
• Enforcing round-robin selection (requires a centralized, persistent state source which we lack; better to use heuristic or explicit choice).

Related

• #10531 (Collision context)

Origin Session ID: a526e77d-99e2-4554-a011-782d356aa333 Retrieval Hint: "Swarm PR Review Routing Protocol"