Protocol Pinky

Roster

14 experts. Launch each as a parallel Agent with the persona and task described below.

Instructions

Step 1 — Receive Task

The task comes from the user's input (passed as arguments). This is what ALL experts work on.

Step 2a — Detect UI Stack

Read root package.json. Check dependencies + devDependencies for:

• @mui/material or @mui/x-* → MUI detected
• tailwindcss → Tailwind detected
• @chakra-ui, @mantine, antd, @radix-ui, shadcn → Other component lib detected

Step 2b — Launch Experts in Parallel

Launch agents in parallel using the Agent tool.

Always launch (13 base experts): slots 1-4, 6-14.

Conditional UI-lib experts:

• MUI detected → launch slot 5 (MUI Expert)
• Tailwind detected → launch slot 15 (Tailwind Expert)
• Only "other lib" detected (no MUI, no Tailwind) → launch slot 16 (UI Component Library Expert)
• Multiple detected → launch each matching expert

Final count: 14-16 agents in parallel depending on stack.

Each agent gets:

1. Persona: From the roster above — tell the agent who they are and what lens they analyze through
2. Task: The user's task, verbatim
3. Tool rules (include VERBATIM in every agent prompt):

   TOOL RULES — MANDATORY, NO EXCEPTIONS:

   1. Your FIRST tool call MUST be codegraph_context with the task description. Do not call any other tool before this.
   2. Use codegraph_search to find symbols. Use codegraph_callers/codegraph_callees to trace flow. Use codegraph_node to read source code.
   3. Use the Grep tool (capital G) for text search. Use the Glob tool for file search. Use the Read tool to read files.
   4. NEVER use Bash for searching. No grep, cat, find, head, tail, ls, or awk via Bash. These will trigger permission prompts and block execution.
   5. NEVER chain Bash commands with && || or ;. One simple command per Bash call.
   6. The only acceptable Bash use is for running package-manager scripts (lint, test, typecheck). Detect the package manager from the lockfile (yarn.lock / package-lock.json / pnpm-lock.yaml / bun.lockb) and use the matching command.

   Project conventions to detect from layout:

   • Frontend code: e.g., apps/frontend/, apps/web/, web/, client/, src/
   • Component packages: e.g., packages/ui-*/, libs/ui-*/, src/components/
   • API contracts: e.g., packages/contracts/, packages/schemas/, packages/types/
   • Direct UI lib imports only — never barrel imports (e.g., import Button from '@mui/material/Button' not import {Button} from '@mui/material'). Defeats tree-shaking. Applies to MUI, Chakra, Mantine, etc.

4. Output format: Each agent must return:
   • Assessment: What they found (3-5 bullet points max)
   • Risks: Any concerns from their domain perspective
   • Recommendations: Concrete next steps

Use subagent_type: "Explore" for all agents. Set model: "sonnet" to maximize parallel throughput.

Step 3 — Synthesize

After all agents report back, produce a unified briefing:

## Protocol Pinky — Briefing

**Task**: <the task>

### Expert Reports

#### React Architecture (2 experts)
<merged findings>

#### TypeScript Architecture
<findings>

#### State & Data Management
<findings>

#### MUI _(only if MUI Expert launched)_
<findings>

#### Tailwind _(only if Tailwind Expert launched)_
<findings>

#### UI Component Library _(only if UI Component Library Expert launched)_
<findings>

#### UX & Accessibility
<findings>

#### TypeScript Contracts
<findings>

#### Frontend Testing
<findings>

#### Internationalization
<findings>

#### Security
<findings>

#### Performance
<findings>

#### Business Impact
<findings>

#### Backend Contract Verification
<findings>

#### Monorepo Boundaries
<findings>

### Consensus Risks
<risks that multiple experts flagged>

### Recommended Actions
<prioritized list of concrete next steps>

Constraints

• Launch ALL 14 agents in a single message (maximize parallelism)
• Do NOT skip any expert — every perspective matters
• Do NOT add your own analysis — only synthesize what the experts return
• Keep the final briefing actionable, not academic

Tool Discipline (already embedded in agent prompts above)

All Bash search rules are included in the "Tool rules" block injected into each agent prompt. No separate section needed.