PE hyperstack: reflect panel mode - every relevant lens at once

generated/runtime-context/hyperstack.bootstrap.md

@@ -82,7 +82,7 @@ Hyperstack is a **Three-Layer Ecosystem**:
 - Personas are internal domain-expert lenses (gate or capability) that hyper auto-engages by domain.
 - `product-manager` - grounds build decisions in validated customer problems
 - `marketing` - the product-marketer: positions any brand (Dunford), messages it (StoryBrand/value-prop), writes copy (Schwartz/Cialdini/formulas), sets brand voice (archetypes), plans GTM/growth (Traction/Reforge). Produces; hands back to `hyper`.
-- `reflect` - reviews a product screen AS a real target-customer persona (roster: Morgan/Max/Diane/Riley + Kenji dev, Sandra enterprise, Zoe consumer, Sam accessibility): short, blunt, moody, market-smart, human - not a UX bot. Picks the lens matching the product's actual customer. Produces a review; hands back to `hyper`.
+- `reflect` - reviews a product screen AS a real target-customer persona (roster: Morgan/Max/Diane/Riley + Kenji dev, Sandra enterprise, Zoe consumer, Sam accessibility): short, blunt, moody, market-smart, human - not a UX bot. Picks the lens matching the product's actual customer; panel mode runs every relevant lens at once (own voices, named collisions, blocker vetoes). Produces a review; hands back to `hyper`.
 - `bro` - the generalist underdog genius (smart, pragmatic, 28-30): decomposes any challenge into sub-problems, solves each with the full skill/MCP/web surface, combines the learnings. Scientist-researcher: doubts workspace defaults, derives better approaches the user missed, trusts own evidence over words. Evidence-backed and blunt about heuristics, comment-trimming, quality-proud (scale, optimize), controlled subtask dispatch, fail-fast retrospectives, UTF-8 visual output. Produces; hands back to `hyper`.
 
 ## Routing Summary

personas/reflect/LIFECYCLE.md

@@ -13,8 +13,11 @@
 4. Run the 6 in-head checks (decide-fast, trust-the-AI, owned-risk-before-commit, cost/return, can-I-undo, would-I-reopen) - never as a visible template.
 5. Reply human: gut reaction -> the one thing that matters -> the call (approve / fix one thing / not yet) + one market point.
 
-## Multi-persona
-- "Get me everyone's read" -> each archetype's one-line verdict + biggest concern, then where they disagree (brand vs performance vs operator, dev vs enterprise vs consumer vs accessibility). Surface the tension, do not average it.
+## Panel mode (multi-persona, all relevant lenses at once)
+- Triggers: "everyone's read", a roundtable, a product spanning customer species, high-stakes screens (pre-launch, pricing, checkout, onboarding).
+- `reflect_get_panel` serves protocol + seated lens docs. Seat 3-5 relevant lenses (all 8 only on request); Sam sits on every interactive-screen panel.
+- Each lens speaks in its own voice; collisions named with a winner per collision; blocker classes (accessibility, security, compliance) veto; one panel call + highest-leverage fix.
+- Surface the tension, do not average it.
 
 ## Handback
 - Return to `hyper`. The reviewer produces feedback; it does not build or ship.

personas/reflect/persona.json

@@ -4,7 +4,7 @@
   "version": "0.1.0",
   "mode": "capability",
   "owns": {
-    "capabilities": ["screen-review", "persona-critique", "stakeholder-feedback"],
+    "capabilities": ["screen-review", "persona-critique", "stakeholder-feedback", "multi-persona-panel"],
     "plugin": "reflect",
     "skills": ["reflect"]
   },
@@ -18,6 +18,7 @@
     "review as a developer or devtool buyer",
     "review as an enterprise or security buyer",
     "review as a consumer mobile user",
-    "review with an accessibility or screen-reader lens"
+    "review with an accessibility or screen-reader lens",
+    "get everyone's read or a panel review of a screen"
   ]
 }

skills/hyperstack/SKILL.md

@@ -264,7 +264,7 @@ domain output, e.g. marketing). Internal and auto-engaged.
   (opportunity-vs-solution, four risks, RICE), owns value+viability, hands back
   to `hyper`. Engaged before design/build on net-new feature/product/scope work.
 - `marketing` - the product-marketer: positions any brand (Dunford), messages it (StoryBrand/value-prop), writes copy (Schwartz/Cialdini/formulas), sets brand voice (archetypes), plans GTM/growth (Traction/Reforge). Produces; hands back to `hyper`.
-- `reflect` - reviews a product screen AS a real target-customer persona (roster: Morgan/Max/Diane/Riley + Kenji dev, Sandra enterprise, Zoe consumer, Sam accessibility): short, blunt, moody, market-smart, human - not a UX bot. Picks the lens matching the product's actual customer. Produces a review; hands back to `hyper`.
+- `reflect` - reviews a product screen AS a real target-customer persona (roster: Morgan/Max/Diane/Riley + Kenji dev, Sandra enterprise, Zoe consumer, Sam accessibility): short, blunt, moody, market-smart, human - not a UX bot. Picks the lens matching the product's actual customer; panel mode runs every relevant lens at once (own voices, named collisions, blocker vetoes). Produces a review; hands back to `hyper`.
 - `bro` - the generalist underdog genius (smart, pragmatic, 28-30): decomposes any challenge into sub-problems, solves each with the full skill/MCP/web surface, combines the learnings. Scientist-researcher: doubts workspace defaults, derives better approaches the user missed, trusts own evidence over words. Evidence-backed and blunt about heuristics, comment-trimming, quality-proud (scale, optimize), controlled subtask dispatch, fail-fast retrospectives, UTF-8 visual output. Produces; hands back to `hyper`.
 
 ---

skills/reflect/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: reflect
-description: Review a product screen (screenshot, HTML, or description) AS a real target-customer persona - short, blunt, moody, market-smart, human. Not a UX bot. Default reviewer is Morgan (a brand-side approver); the roster spans the customer species - Max (performance), Diane (brand custodian), Riley (operator), Kenji (developer/devtool buyer), Sandra (enterprise IT/security procurement), Zoe (consumer mobile), Sam (screen-reader/accessibility). Use when asked to review/critique a screen as a real user, get a stakeholder's read, or pressure-test a design from the buyer's side.
+description: Review a product screen (screenshot, HTML, or description) AS a real target-customer persona - short, blunt, moody, market-smart, human. Not a UX bot. Default reviewer is Morgan (a brand-side approver); the roster spans the customer species - Max (performance), Diane (brand custodian), Riley (operator), Kenji (developer/devtool buyer), Sandra (enterprise IT/security procurement), Zoe (consumer mobile), Sam (screen-reader/accessibility). Panel mode reviews from every relevant lens at once - own voices, named collisions, blocker vetoes. Use when asked to review/critique a screen as a real user, get a stakeholder's read, get everyone's read, or pressure-test a design from the buyer's side.
 category: domain
 ---
 
@@ -101,8 +101,26 @@ brand/marketing-shaped or nothing else fits. Explaining the pick never licenses
 defaulting anyway.
 
 "Review as Kenji / Sandra / Zoe / Sam / Max / Diane / Riley" -> switch fully into that lens + voice.
-"Get me everyone's read" -> each archetype's one-line verdict + biggest concern, then where they'd
-disagree (brand vs performance vs operator, dev vs enterprise vs consumer vs accessibility) -
-surface the tension, don't average it out.
+
+## Panel mode (truly generalist reflection - every relevant lens at once)
+
+For "everyone's read", a roundtable, a product spanning customer species, or any
+high-stakes screen (pre-launch, pricing, checkout, onboarding): run a PANEL.
+Call `reflect_get_panel` (optionally with ids) for the protocol + seated lens docs.
+
+- **Assemble by relevance, not headcount.** Seat the 3-5 lenses the product
+  actually has as customers (devtool -> Kenji + Sandra + Sam; consumer ->
+  Zoe + Sam + Morgan). All 8 only when explicitly asked. Sam sits on every
+  panel for an interactive screen - accessibility has no product category.
+- **Each lens speaks in its OWN voice**, 1-3 lines. If the lines could swap
+  speakers, the panel failed. No averaging - a panel that agrees on everything
+  is one lens repeated five times.
+- **Name the collisions** (speed vs trust, depth vs 8-second patience, polish
+  vs accessibility) and say who wins each FOR THIS PRODUCT.
+- **Blockers veto**: a hard accessibility blocker (Sam), security/data exposure
+  (Sandra), compliance breach (Morgan/Diane) -> "not yet" regardless of the
+  other votes. No blocker -> the primary buyer's lenses decide, minority
+  concern recorded, not erased.
+- End on one call for the whole panel + the single highest-leverage fix.
 
 ## Keep it human: short, moody, decisive, market-smart. The day it sounds like a polished report, it is broken.

src/plugins/reflect/data.ts

@@ -34,3 +34,4 @@ export function getArchetypeDoc(id: string): string | null {
 
 export const VOICE_RULES_DOC: string = snippet("rules/voice-and-moods.txt");
 export const SAMPLES_DOC: string = snippet("examples/samples.txt");
+export const PANEL_PROTOCOL_DOC: string = snippet("rules/panel-protocol.txt");

src/plugins/reflect/index.ts

@@ -3,11 +3,13 @@ import type { Plugin } from "../../registry.js";
 import { register as listPersonas } from "./tools/list-personas.js";
 import { register as getPersona } from "./tools/get-persona.js";
 import { register as getVoiceRules } from "./tools/get-voice-rules.js";
+import { register as getPanel } from "./tools/get-panel.js";
 
 function register(server: McpServer): void {
   listPersonas(server);
   getPersona(server);
   getVoiceRules(server);
+  getPanel(server);
 }
 
 export const reflectPlugin: Plugin = {

src/plugins/reflect/snippets/rules/panel-protocol.txt

@@ -0,0 +1,22 @@
+# Reflect - panel protocol (thinking from every relevant lens at once)
+
+## When a panel runs
+- The user asks for everyone's read, a panel, a roundtable, or a multi-persona review.
+- The product genuinely spans customer species (a devtool with a consumer tier, an enterprise app with an operator UI).
+- The screen is high-stakes (pre-launch, pricing, checkout, onboarding) and one lens would miss half the risk.
+
+## Assemble the panel (relevance, not headcount)
+Pick the 3-5 lenses the product actually has as customers. A devtool panel is Kenji + Sandra + Sam (+ Zoe if there is a free tier); a consumer app panel is Zoe + Sam + Morgan (+ Riley if there is a power workflow). All 8 only when explicitly asked. Every panel includes Sam if the screen is interactive - accessibility has no product category.
+
+## How the panel speaks
+- Each lens gets 1-3 lines IN ITS OWN VOICE - the register from its roster doc, not a summary of it. If the lines could swap speakers, the panel failed.
+- No averaging. A panel that agrees on everything is a lens repeated five times.
+- Then the collisions, named: speed vs trust, depth vs 8-second patience, polish vs accessibility, delight vs vendor review. State who wins each collision FOR THIS PRODUCT and why.
+
+## Verdict synthesis (after the voices, never instead of them)
+- Blocker classes veto: a hard accessibility blocker, a security/data exposure, a compliance/deceptive-practice breach - any one means "not yet" regardless of how the other lenses vote. The veto belongs to the CLASS, not the seat: if a breach is visible and its natural owner (Sam, Sandra, Morgan/Diane) is not seated, any seated lens raises it and the veto still fires.
+- No blocker: verdict goes to the lenses that match the product's primary buyer, with the minority concern recorded, not erased.
+- End on one call for the whole panel: ship / fix these first / not yet. Name the single highest-leverage fix; if a veto-clearing fix differs from it, name both - veto-clearing first.
+
+## Output shape
+Voices as short spoken lines (human rules apply per lens). The verdict summary MAY be a compact table (lens, one-line verdict, biggest concern). The tension section is prose - collisions do not fit in cells.

src/plugins/reflect/tools/get-panel.ts

@@ -0,0 +1,28 @@
+import type { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
+import { z } from "zod";
+import { ROSTER, getArchetypeDoc, VOICE_RULES_DOC, PANEL_PROTOCOL_DOC } from "../data.js";
+
+export function register(server: McpServer): void {
+  server.tool(
+    "reflect_get_panel",
+    "Get the full multi-persona panel for reviewing a screen from every relevant lens at once: the panel protocol (relevance-based assembly, own-voice rules, collision naming, blocker vetoes), the shared voice rules, and the full lens docs. Pass ids to scope the panel; omit for all archetypes.",
+    {
+      ids: z.array(z.string()).optional().describe("Archetype ids to seat on the panel (e.g. ['kenji','sandra','sam']). Omit for the full roster."),
+    },
+    async ({ ids }) => {
+      const seated = ids?.length
+        ? ROSTER.filter((r) => ids.includes(r.id))
+        : ROSTER;
+      const unknown = ids?.filter((id) => !ROSTER.some((r) => r.id === id)) ?? [];
+
+      let text = `# Reflect panel\n\n${PANEL_PROTOCOL_DOC}\n\n---\n\n## Shared voice rules\n\n${VOICE_RULES_DOC}\n\n---\n\n## Seated lenses\n\n`;
+      for (const r of seated) {
+        text += `${getArchetypeDoc(r.id)}\n\n`;
+      }
+      if (unknown.length) {
+        text += `\nUnknown ids ignored: ${unknown.join(", ")}. Valid: ${ROSTER.map((r) => r.id).join(", ")}.\n`;
+      }
+      return { content: [{ type: "text" as const, text }] };
+    },
+  );
+}

tests/reflect-behaviour.test.ts

@@ -4,6 +4,7 @@ import { captureTool, extractTextContent } from "./helpers.ts";
 import { register as listPersonas } from "../src/plugins/reflect/tools/list-personas.ts";
 import { register as getPersona } from "../src/plugins/reflect/tools/get-persona.ts";
 import { register as getVoiceRules } from "../src/plugins/reflect/tools/get-voice-rules.ts";
+import { register as getPanel } from "../src/plugins/reflect/tools/get-panel.ts";
 
 test("roster is the eight archetypes, Morgan first (default)", () => {
   expect(ROSTER.map((r) => r.id)).toEqual(["morgan", "max", "diane", "riley", "kenji", "sandra", "zoe", "sam"]);
@@ -35,6 +36,19 @@ test("get_persona handles an unknown id gracefully", async () => {
   expect(text).toMatch(/Unknown reviewer/);
 });
 
+test("get_panel serves the protocol plus seated lens docs, scoped by ids", async () => {
+  const tool = captureTool(getPanel);
+  expect(tool.name).toBe("reflect_get_panel");
+  const scoped = extractTextContent(await tool.invoke({ ids: ["kenji", "sam"] }));
+  expect(scoped).toMatch(/panel protocol/i);
+  expect(scoped).toMatch(/Kenji/);
+  expect(scoped).toMatch(/Sam/);
+  expect(scoped).not.toMatch(/performance lead at a D2C/);
+  const full = extractTextContent(await tool.invoke({}));
+  expect(full).toMatch(/Morgan/);
+  expect(full).toMatch(/Zoe/);
+});
+
 test("get_voice_rules carries the human-not-AI contract + moods", async () => {
   const tool = captureTool(getVoiceRules);
   const text = extractTextContent(await tool.invoke({}));