Relay parallel agents

Skill Genb2nu/Relay/skills/relay-parallel-agents

Patterns for dispatching Relay agents in parallel. Covers Phase 3 (Auditor + Warden), Phase 5 (Vault + Stylist), and Phase 6 (Sentinel + Warden). Both agents must complete and pass before the phase gate clears. Adapted from Superpowers dispatching-parallel-agents — no external dependency required.From its SKILL.md

Install

npx -y skills add Genb2nu/Relay --skill relay-parallel-agents

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.7 KB, as published. Nobody here has run it

Relay Parallel Agents

Adapted from Superpowers dispatching-parallel-agents for Power Platform review workflows.

When Relay Uses Parallel Agents

Phase	Agents	Both must pass
Phase 3 — Review	Auditor + Warden	Yes — both must approve plan
Phase 5 — Build	Vault + Stylist	Yes — both outputs needed before Forge specialists
Phase 6 — Verify	Sentinel + Warden	Yes — both must pass before ship

Core Principle

Parallel agents are independent specialists reviewing or building the same artefacts from different perspectives. They must never see each other's output before forming their own judgment — this preserves independence.

❌ Wrong: "Auditor found 3 issues. Now Warden, review the plan knowing this."
✅ Right: Dispatch Warden with only plan.md + security-design.md, not Auditor's report.

Phase 3 — Auditor + Warden Parallel Review

Both receive the same input: docs/plan.md + docs/security-design.md

Dispatch Auditor with:

plan.md (full content)
requirements.md (for cross-checking completeness)
Auditor persona rules

Dispatch Warden with:

plan.md (full content)
security-design.md (full content)
Warden persona rules
skills/power-platform-security-patterns/SKILL.md

After both complete:

Collect all issues from both reports
Drafter addresses ALL issues from BOTH agents
Re-dispatch both for re-review
Loop until BOTH give explicit approval
Only then proceed to Phase 4

Do NOT proceed if:

Only one agent approved
Either agent gave conditional approval without re-review
Issues were "noted" but not fixed

Phase 5 — Vault + Stylist Parallel Build

These are parallel BUILD tasks, not reviews. Both produce outputs Forge needs.

Vault produces:

Dataverse schema (tables, columns, relationships, choices)
Security roles + FLS profiles
Environment variables
Seed data
Updates state.json with component GUIDs

Stylist produces:

docs/design-system.md (colours, typography, spacing, component patterns)

Forge specialists wait for both before starting builds. forge-canvas reads design-system.md to inform /generate-canvas-app prompt. All Forge specialists read state.json to find Vault's component GUIDs before creating anything.

If Stylist is unavailable: forge-canvas proceeds with default design tokens but flags Canvas App as needing visual review.

Phase 6 — Sentinel + Warden Parallel Verify

Both verify the completed build from different angles.

Dispatch Sentinel with:

docs/plan.md (the approved spec)
List of all components built (from Forge specialist/Vault handoffs)
Verification persona rules

Sentinel checks: does what was built match what was planned?

Dispatch Warden with:

docs/security-design.md (the approved security spec)
List of all components built
Warden persona rules

Warden checks: are the security controls actually in place in the built solution?

After both complete:

If both pass → proceed to Phase 7 (wrap-up + export)
If either fails → Forge specialist/Vault fix the specific failures → re-verify
Loop until both pass

Context Isolation

Each agent receives a precisely crafted context — only what it needs:

Auditor context:
  - plan.md (the document being reviewed)
  - requirements.md (to check completeness)
  - auditor persona rules

Warden context:
  - plan.md (for architecture)
  - security-design.md (the document being reviewed)
  - security-patterns skill
  - warden persona rules

Neither agent sees the other's output before completing their own review.

Copilot VS Code — Sequential Fallback

Copilot VS Code doesn't support true parallel subagents or /fleet. Run Phase 3 as: Auditor first, then Warden, both in same session. Conductor still collects both outputs before proceeding. Do NOT let Warden see Auditor's output before completing its own review.

Important: /fleet is exclusive to Copilot CLI. In VS Code Agent Mode, Conductor runs all agents sequentially with identical output quality. Do not attempt to use /fleet syntax in VS Code — it will be ignored. The sequential approach is functionally equivalent; only parallelism (speed) differs.