@hegemonart/get-design-done 1.36.2 → 1.36.3

package/.claude-plugin/marketplace.json

@@ -5,14 +5,14 @@
   },
   "metadata": {
     "description": "Get Design Done — 5-stage agent-orchestrated design pipeline with 9 connections, handoff-first workflow, bidirectional Figma write-back, 22+ specialized agents, queryable knowledge layer (intel store, dependency analysis, learnings extraction), and a self-improvement loop (reflector, frontmatter + budget feedback, global-skills layer). v1.20.0 ships the SDK foundation: gdd-state MCP server (11 typed tools), lockfile-safe STATE.md mutations, event stream, and resilience primitives (jittered-backoff, rate-guard, error-classifier, iteration-budget) for rate-limit + 429 + context-overflow recovery. Full CI/CD pipeline (Node 22/24 × Linux/macOS/Windows) and release automation (auto-tag + GitHub Release + release-time smoke test).",
-    "version": "1.36.2"
+    "version": "1.36.3"
   },
   "plugins": [
     {
       "name": "get-design-done",
       "source": "./",
       "description": "Agent-orchestrated 5-stage design pipeline: Brief → Explore → Plan → Design → Verify. 22+ specialized agents, 9 connections (Figma, Refero, Preview, Storybook, Chromatic, Figma Writer, Graphify, Pinterest, Claude Design), Claude Design handoff, bidirectional Figma write-back, and a queryable intel store (.design/intel/) for dependency and learnings queries. Standalone commands: style, darkmode, compare, figma-write, graphify, handoff, analyze-dependencies, skill-manifest, extract-learnings. Embeds NNG heuristics, WCAG thresholds, typographic systems, motion framework, and anti-pattern catalog. Ships with a full CI/CD pipeline (Node 22/24 × Linux/macOS/Windows) and release automation. Optimization layer (v1.0.4.1, retroactive): gdd-router + gdd-cache-manager skills, PreToolUse budget-enforcer hook, tier-aware agent frontmatter, lazy checker gates, streaming synthesizer, /gdd:warm-cache + /gdd:optimize commands, and cost telemetry at .design/telemetry/costs.jsonl — targeting 50-70% per-task token-cost reduction with no quality-floor regression. v1.20.0 SDK foundation: gdd-state MCP server (11 typed tools), lockfile-safe STATE.md mutations, event stream at .design/telemetry/events.jsonl, resilience primitives (jittered-backoff, rate-guard, error-classifier, iteration-budget) with rate-limit + 429 + context-overflow recovery, and TypeScript toolchain.",
-      "version": "1.36.2",
+      "version": "1.36.3",
       "author": {
         "name": "hegemonart"
       },

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "get-design-done",
   "short_name": "gdd",
-  "version": "1.36.2",
+  "version": "1.36.3",
   "description": "Agent-orchestrated 5-stage design pipeline: Brief → Explore → Plan → Design → Verify. 22+ specialized agents, 9 connections (Figma, Refero, Preview, Storybook, Chromatic, Figma Writer, Graphify, Pinterest, Claude Design), handoff-first workflow via Claude Design bundles, bidirectional Figma write-back (annotations, Code Connect), queryable intel store (`.design/intel/`) for O(1) design surface lookups, and self-improvement loop (reflector agent, frontmatter + budget feedback, global-skills layer at `~/.claude/gdd/global-skills/`). Standalone commands: style, darkmode, compare, figma-write, graphify, handoff, analyze-dependencies, skill-manifest, extract-learnings, reflect, apply-reflections. Embeds NNG heuristics, WCAG thresholds, typographic systems, motion framework, and anti-pattern catalog. Ships with a full CI/CD pipeline (Node 22/24 × Linux/macOS/Windows, lint + schema + frontmatter + stale-ref + shellcheck + gitleaks + injection-scan + blocking size-budget) and release automation (auto-tag + GitHub Release + release-time smoke test). Optimization layer (v1.0.4.1, retroactive): gdd-router + gdd-cache-manager skills, PreToolUse budget-enforcer hook, tier-aware agent frontmatter, lazy checker gates, streaming synthesizer, /gdd:warm-cache + /gdd:optimize commands, and cost telemetry at .design/telemetry/costs.jsonl — targeting 50-70% per-task token-cost reduction with no quality-floor regression. v1.20.0 SDK foundation: gdd-state MCP server (11 typed tools), lockfile-safe STATE.md mutations, event stream at .design/telemetry/events.jsonl, resilience primitives (jittered-backoff, rate-guard, error-classifier, iteration-budget) with rate-limit + 429 + context-overflow recovery, and TypeScript toolchain. v1.27.7 ships gdd-mcp (Phase 27.7): 12 read-only MCP tools for sub-3s priming. v1.28.0 (Phase 28): Foundational References Tier 2 — 5 new reference files (color-theory, composition, proportion-systems, i18n, contrast-advanced), 2 verifier i18n probes + 1 explore i18n-readiness probe, 12 additive cross-link insertions across 10 existing references, 2 orthogonal audit-scoring lens-tags (composition_alignment + i18n_readiness).",
   "author": {
     "name": "hegemonart",

package/CHANGELOG.md

@@ -4,6 +4,26 @@ All notable changes to get-design-done are documented here. Versions follow [sem
 
 ---
 
+## [1.36.3] - 2026-06-01
+
+### Phase 36.3 — Knowledge Tier-3: Conversational UI — completes Phase 36
+
+Third and **FINAL** sub-phase of the split **Phase 36 (Knowledge Tier 3)** — completing it marks the **parent Phase 36 COMPLETE** (domain packs 36.1 + motion-tool verification 36.2 + conversational UI 36.3 all shipped). Conversational UI was zero-coverage even though it's a real surface (chatbot empty-states, voice flow design, prompt-as-UX). **No new pillar, no new runtime dependency, no new egress** (reference markdown + an agent-prompt enum addition).
+
+### Added
+
+- **`reference/conversational-ui.md`** — voice-flow patterns (no-input / no-match reprompts, confirmation, human handoff), multi-turn dialogue (context carryover, slot-filling, repair), prompt-as-UX (the assistant persona/tone/boundaries as a versioned design artifact), chatbot empty-states + suggested replies, voice-first onboarding, and error recovery + accessibility (transcripts/captions). Carries a `## Detection signals` section + an `## Audit checklist`. Registered in `reference/registry.json` (`type: heuristic`, `phase: 36.3`; 149 → 150). CLI/REPL UX is out of scope this phase.
+- **`agents/design-context-builder.md`** — Step 0E gains a **7th** project type `conversational` (enum 6 → 7). It routes to `design-executor` (a chat widget / voice-app card is still rendered code) **and loads `reference/conversational-ui.md`** for the interaction patterns. Detection from brief keywords (chatbot, voice, assistant, conversational) + `package.json` deps (`botpress`, `rasa`, `dialogflow`, `actions-on-google`, `ask-sdk-core`, `botframework`). The first six project types remain the Phase-34 rendered-output set; `conversational` is the Phase-36.3 interaction-surface type on the same axis.
+
+### Notes
+
+- **No new runtime dependency, no new egress.** Detection is the agent matching keywords/deps; no code, no network.
+- 6-manifest lockstep at **v1.36.3** + `OFF_CADENCE_VERSIONS.add('1.36.3')` + the 25 live-pinned `manifests-version.txt` baselines forward-propagated 1.36.2 → 1.36.3.
+- Inventory relock: registry-diff 149 → 150, tarball golden 669 → 670 (+`reference/conversational-ui.md`). No new connection/skill/agent dir. `design-verifier.md` deliberately untouched (it is at its 700-line cap; conversational verify rides the normal web path).
+- **This completes the parent Phase 36 (Knowledge Tier 3) — domain packs + motion-tool verification + conversational UI.**
+
+---
+
 ## [1.36.2] - 2026-06-01
 
 ### Phase 36.2 — Knowledge Tier-3: Motion-Tool Verification (Lottie + Rive)

package/README.md

@@ -158,6 +158,10 @@ Four industry-specific design-pattern packs at [`reference/domains/`](reference/
 
 GDD now opens the motion **exports** a project ships. The pure, dependency-free [`validate-motion`](scripts/lib/motion/validate-motion.cjs) checks a **Lottie** JSON for frame-rate sanity, non-positive duration, embedded-asset bloat, and a perf budget (`MO-*` rules); for **Rive** `.riv` (binary) it does size + a `RIVE`-header sanity, and surfaces the state-machine graph (unreachable states, no-exit loops) when the opt-in Rive runtime is present. The [`motion-verifier`](agents/motion-verifier.md) agent runs at verify time ([`connections/lottie.md`](connections/lottie.md) / [`connections/rive.md`](connections/rive.md)) and **WARNs — never blocks** (motion is creative). **No new runtime dependency** — the Lottie player and Rive runtime are opt-in. Second sub-phase of Phase 36.
 
+### Conversational UI (v1.36.3)
+
+[`reference/conversational-ui.md`](reference/conversational-ui.md) closes a zero-coverage surface: voice flows + chatbots. It codifies voice-flow reprompts (no-input / no-match → human handoff), multi-turn dialogue (context carryover, slot-filling, repair), **prompt-as-UX** (the assistant's persona/tone/boundaries as a versioned design artifact), chatbot empty-states + suggested replies, voice-first onboarding, and error recovery + accessibility. [`design-context-builder`](agents/design-context-builder.md) detects a `conversational` project type (brief keywords + chatbot/voice deps like `botpress` / `dialogflow` / `actions-on-google` / `ask-sdk-core`) and loads the pack into context. **No new runtime dependency.** This is the **final sub-phase of Phase 36 — which is now complete** (domain packs + motion-tool verification + conversational UI).
+
 ### Previous releases
 
 - **v1.26.0** — Headless Model Resolver (per-runtime tier→model map, `resolved_models` router field, per-runtime price tables, `reasoning-class` runtime-neutral alias).

package/agents/design-context-builder.md

@@ -218,7 +218,7 @@ Proceed to Step 0E regardless of whether Step 0D ran or was skipped.
 
 Detect the **project type** so the pipeline routes the brief to the correct executor. Reuse the Step 0C / Step 1 grep/glob idiom (file reads only, < 1 second, no skip condition).
 
-**Enum (6 values — D-06, the full set):** `web` (DEFAULT) · `native-ios` · `native-android` · `flutter` · `email` · `print`.
+**Enum (7 values — D-06 + 36.3):** `web` (DEFAULT) · `native-ios` · `native-android` · `flutter` · `email` · `print` · `conversational`. (The first six are the Phase-34 *rendered-output* set; `conversational` is the Phase-36.3 *interaction-surface* type — a chat/voice UI is still rendered code, so it routes to `design-executor` but loads the conversational patterns.)
 
 **Detection signals + precedence** (first match wins; brief overrides — if the user explicitly says "iOS app" / "Android app" / "Flutter app" / "email" / "newsletter" / "email template" / "print" / "PDF" / "print-ready" / "brochure" / "flyer" / "poster", honor that):
 
@@ -228,10 +228,11 @@ ls *.xcodeproj Package.swift 2>/dev/null          # → native-ios   (when no pu
 ls build.gradle build.gradle.kts settings.gradle 2>/dev/null  # → native-android (when no pubspec)
 ls **/*.mjml email/ emails/ templates/email 2>/dev/null  # → email (email-template signals)
 ls **/*.print.css print/ templates/print 2>/dev/null  # → print (print-output signals: a print stylesheet / print-template dir)
+grep -lE '"(botpress|@botpress/|rasa|dialogflow|@google-cloud/dialogflow|actions-on-google|ask-sdk-core|botframework-webchat|@microsoft/botframework)"' package.json 2>/dev/null  # → conversational (chatbot/voice-assistant deps)
 ls package.json 2>/dev/null                       # → web          (default; also the fallback when none match)
 ```
 
-Precedence: an explicit brief override (the user says "email" / "newsletter" / "email template", or "print" / "PDF" / "print-ready" / "brochure" / "flyer" / "poster") wins like the other brief-overrides; otherwise `pubspec.yaml` (flutter) > `*.xcodeproj`/`Package.swift` (native-ios) > `build.gradle*`/`settings.gradle` (native-android) > `.mjml` files / an `email/` templates directory (email) > a print stylesheet (`*.print.css`) / a `print/` templates directory (print) > `package.json` / none (web — DEFAULT).
+Precedence: an explicit brief override (the user says "email" / "newsletter" / "email template", or "print" / "PDF" / "print-ready" / "brochure" / "flyer" / "poster", or "chatbot" / "voice app" / "Alexa skill" / "conversational" / "assistant") wins like the other brief-overrides; otherwise `pubspec.yaml` (flutter) > `*.xcodeproj`/`Package.swift` (native-ios) > `build.gradle*`/`settings.gradle` (native-android) > `.mjml` files / an `email/` templates directory (email) > a print stylesheet (`*.print.css`) / a `print/` templates directory (print) > chatbot/voice deps (`botpress`/`rasa`/`dialogflow`/`actions-on-google`/`ask-sdk-core`/`botframework`) (conversational) > `package.json` / none (web — DEFAULT).
 
 **Routing table** (project type → executor — one row per type, trivially appendable):
 
@@ -243,10 +244,11 @@ Precedence: an explicit brief override (the user says "email" / "newsletter" / "
 | flutter          | flutter-executor  |
 | email            | email-executor    |
 | print            | pdf-executor      |
+| conversational   | design-executor (loads `reference/conversational-ui.md`) |
 
 <!-- Phase 34 output types complete: native (34.1: native-ios/native-android/flutter) + email (34.2) + print (34.3). Print is the FINAL Phase-34 output type — no further Phase-34 output types; the enum + routing table above are the full set. (34.1/34.2 kept this seam OPEN for the next type; 34.3 ties it off.) -->
 
-Record the detected type in DESIGN-CONTEXT.md as a `<project_type>` line (e.g. `<project_type>native-ios</project_type>`) so downstream stages route correctly. The native specifics (token→theme bridge) live in `reference/native-platforms.md`; the email specifics (table layout, inline styles, MSO/dark-mode constraints) live in `reference/email-design.md`; the print specifics (the `@page` box model, bleed/crop marks, CMYK awareness, font embedding, 300dpi raster) live in `reference/print-design.md` — do not inline any of them here.
+Record the detected type in DESIGN-CONTEXT.md as a `<project_type>` line (e.g. `<project_type>native-ios</project_type>`) so downstream stages route correctly. The native specifics (token→theme bridge) live in `reference/native-platforms.md`; the email specifics (table layout, inline styles, MSO/dark-mode constraints) live in `reference/email-design.md`; the print specifics (the `@page` box model, bleed/crop marks, CMYK awareness, font embedding, 300dpi raster) live in `reference/print-design.md`; the conversational specifics (voice-flow reprompts, multi-turn dialogue, prompt-as-UX, chatbot empty-states, error recovery) live in `reference/conversational-ui.md` — do not inline any of them here.
 
 Proceed to Step 1 regardless of outcome.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hegemonart/get-design-done",
-  "version": "1.36.2",
+  "version": "1.36.3",
   "description": "A design-quality pipeline for AI coding agents: brief, plan, implement, and verify UI work against your design system.",
   "author": "Hegemon",
   "homepage": "https://github.com/hegemonart/get-design-done",

package/reference/conversational-ui.md

@@ -0,0 +1,110 @@
+# Conversational UI Design Patterns
+
+This pack covers conversational interfaces in two modes: **voice flows** (IVR, smart-speaker skills/actions, voice assistants) and **chatbot / text assistants** (in-app chat, support bots, LLM-backed copilots). GDD's design-context-builder loads this reference when it detects a conversational or voice project so downstream design and audit stages share a vocabulary. CLI / REPL / terminal UX is explicitly out of scope for this pack — those follow a different interaction model and have their own conventions.
+
+The guidance below leans on established authorities: Google's Conversation Design guidelines, the Amazon Alexa Design Guide, and Nielsen Norman Group's research on chatbots and conversational interfaces.
+
+## Voice-flow patterns
+
+A voice flow is more than its happy path. The happy path is the shortest successful route (user states a complete, in-scope request; system confirms and acts). Design quality lives in the branches around it.
+
+| Branch | Trigger | Pattern |
+|--------|---------|---------|
+| No-input (NID) | User says nothing within the listen window | Escalating reprompts: 1st is a light nudge ("Which city?"), 2nd adds an example ("For example, say Berlin."), 3rd offers help or hands off. Never repeat the same words. |
+| No-match (NM) | Speech recognized but no intent matched | Reflect what was heard, narrow the ask, give an example. After repeated NM, escalate to options or a human. |
+| Confirmation | Action is risky, costly, or irreversible | See implicit vs explicit below. |
+| Barge-in | User speaks over the prompt | Allow it — stop TTS immediately and listen. Disable only for legally required disclosures. |
+| Hand-off | N consecutive failures (typically 2-3) | Route to a human, a visual fallback, or a clear "let's try later." Track the failure count across turns. |
+
+**One-breath confirmation.** Confirm in a single short clause that the user can absorb in one breath: "Two tickets for Friday — booking now." Avoid reading back every slot verbatim; that is robotic and slow.
+
+**Implicit vs explicit confirmation.**
+- *Implicit* — fold the understood value into the next prompt ("Friday — what time?"). Use for low-risk, easily-reversed actions. It keeps the dialogue moving.
+- *Explicit* — require a yes/no ("Send $200 to Sam — confirm?"). Reserve for irreversible, costly, or destructive actions (payments, deletes, sends).
+
+**Failure escalation.** Each successive reprompt should add information (an example, then a menu, then a human). Three strikes is a common ceiling — after that, stop asking and offer an exit. A dead end ("Sorry, goodbye.") is a design failure.
+
+## Multi-turn dialogue rules
+
+Multi-turn dialogue is where conversational UIs earn their name. The core obligation: remember context so the user speaks like a human, not a form.
+
+- **Context carryover (anaphora).** Resolve references to prior turns. After "Weather in Paris?", the user says "What about tomorrow?" or "And Rome?" — carry the unstated slots (location, intent) forward. Failing this forces the user to repeat themselves and breaks the illusion of conversation.
+- **Turn-taking.** Make it unambiguous whose turn it is. In voice, an earcon or a brief pause signals "your turn." In chat, a typing indicator signals the assistant is composing. Never leave the user unsure whether to speak/type.
+- **Disambiguation.** When input matches more than one intent or entity, ask a tight either/or ("Did you mean the 9am or the 9pm flight?") rather than restarting. Offer at most 2-3 options by voice.
+- **Slot-filling + validation.** Collect required slots one at a time when missing, but accept them all at once when volunteered ("Book a 7pm table for four" fills time + party size in one turn). Validate each slot against real constraints (date in the future, party size within limits) and reprompt only the invalid slot, not the whole request.
+- **Repair.** Support correction mid-flow: "No, I said *four*, not *forty*." Detect the correction marker ("no", "I meant", "actually") and overwrite the targeted slot without discarding the rest of the context.
+- **Don't over-ask.** Never re-confirm what is already certain. If the user said "tomorrow at noon," do not ask "what day?" again. Over-asking is the most common way a multi-turn flow feels broken.
+
+## Prompt-as-UX (system prompts as design artifacts)
+
+For LLM-backed assistants, the `system-prompt` is a design artifact, not config. It defines the assistant's persona, tone, scope, and boundaries — and those ARE the user experience. Treat it accordingly.
+
+- **Persona and voice are design decisions.** Warmth, formality, verbosity, use of humor, first-person vs neutral — these are chosen, documented, and kept consistent across every turn. A persona that drifts mid-conversation reads as broken.
+- **The prompt is the spec for behavior.** What the assistant will and won't do, how it greets, how it refuses, how it asks for clarification — all of it is specified in the prompt. Downstream copy and flows must match it.
+- **Version and review it like production copy.** The `system-prompt` belongs in source control, with diffs reviewed by design/content the same way UI strings are. A wording change to a refusal message is a UX change.
+- **Refusal and limits messaging.** When the assistant can't or won't help, the decline should be in-voice, specific, and ideally redirect ("I can't process refunds here, but I can show you the refund form."). A blunt "I cannot do that" is a UX bug.
+- **Consistency of voice.** Error states, confirmations, empty states, and refusals must all sound like the same persona. Audit them together, not in isolation.
+
+## Chatbot empty-states and entry
+
+The opening message is the chatbot's most important screen. It does the work an empty form can't: it sets scope and teaches capability.
+
+- **The opener states scope + capability.** "Hi — I can track orders, start a return, or check store hours. What do you need?" tells the user what's in bounds before they guess wrong. Vague openers ("How can I help?") invite out-of-scope requests the bot then fails.
+- **Suggested-reply chips.** Offer 2-4 tappable starter prompts that double as a capability menu. They remove the cold-start guess and demonstrate the input format.
+- **Discoverability of commands.** If the bot supports commands or shortcuts, surface them (a menu, a persistent "Help" affordance, or a slash-command hint). Hidden capabilities don't exist to users.
+- **Avoid the blank-box-no-affordance trap.** A bare text input with a blinking cursor and no guidance is the #1 chatbot empty-state failure (NN/g). Always pair the input with scope-setting copy and/or chips.
+- **Re-entry and history.** On return, remind the user where they left off or what the bot does — don't assume they remember the opener from last session.
+
+## Voice-first onboarding
+
+Voice-first surfaces (smart speakers, headless assistants) onboard without a screen, which makes discoverability the central challenge.
+
+- **Teach invocation.** Make the wake phrase / invocation name explicit early ("Just say 'Ask Acme to…'"). Users can't use what they can't name.
+- **Teach capability by example.** Replace abstract menus with concrete sample utterances: "You can say 'add milk to my list' or 'what's on my list?'" The "you can say…" pattern is the canonical voice-discoverability tool.
+- **Progressive disclosure.** Don't dump every feature in the first session. Reveal capabilities over time, contextually, as the user's needs surface.
+- **Earcons and audio feedback.** Use short, distinct audio cues to mark state — listening, processing, success, error. Audio is the only "UI" the user has; consistent earcons replace visual affordances.
+- **Confirm understanding audibly.** Because there's no screen to glance at, brief spoken confirmations carry the load that visual feedback would in a GUI.
+
+## Error recovery in voice flows
+
+Recovery is the difference between a usable voice product and an abandoned one. The guiding rule: **never a dead end.**
+
+- **Graceful degradation.** When confidence is low, narrow rather than fail — fall back from open prompt to a menu, from a menu to yes/no, from voice to a visual/text fallback if a screen is available.
+- **Rapid, varied reprompts.** Reprompt quickly and change the wording each time. Repeating the identical prompt is the fastest way to frustrate.
+- **Offer alternatives.** "I couldn't find that. Want me to search nearby instead?" keeps momentum and gives the user a next move.
+- **Escalate to human / visual fallback.** After N failed turns, route to a person, send a link to a phone/companion screen, or switch channels. Honor explicit "talk to a person" requests immediately.
+- **Accessibility.** Provide captions and transcripts for voice interactions; ensure chat is screen-reader-friendly with proper roles, focus management, and announced new messages. Don't gate critical actions behind audio-only or visual-only cues — offer both paths.
+
+## Detection signals
+
+The design-context-builder classifies a project as conversational when it sees these signals.
+
+**Keywords (in README, package name, docs, route/handler names):** chatbot, conversational, voice, assistant, dialogue, IVR, utterance, intent, skill, action, slot, prompt, agent, NLU.
+
+**`package.json` dependencies:**
+
+| Dependency | Platform |
+|------------|----------|
+| `botpress` / `@botpress/...` | Botpress conversational platform |
+| `rasa` / `@rasa/...` | Rasa NLU / dialogue (often a Python sibling service) |
+| `dialogflow` / `@google-cloud/dialogflow` | Google Dialogflow NLU |
+| `actions-on-google` | Google Assistant actions |
+| `ask-sdk-core` | Amazon Alexa skills |
+| `botframework-webchat` / `@microsoft/botframework-*` | Microsoft Bot Framework |
+
+Presence of any of these deps is a strong signal; keyword matches add weight. A `system-prompt` / persona file plus an LLM SDK also indicates a text assistant even without the platform deps above.
+
+## Audit checklist
+
+1. Every voice flow handles **no-input (NID)** and **no-match (NM)** with escalating reprompts that add help each time — never a dead end.
+2. Risky, costly, or irreversible actions use **explicit confirmation**; low-risk actions use implicit confirmation and don't over-confirm.
+3. **Barge-in** is allowed on standard prompts (TTS stops and listening resumes immediately), disabled only where legally required.
+4. Voice flows offer a **human / visual fallback after N failed turns** (typically 2-3) and honor explicit "talk to a person" requests.
+5. **Multi-turn context carries over** (anaphora resolved) so the user isn't forced to repeat slots already provided.
+6. **Slot validation** reprompts only the invalid slot and accepts multiple slots when volunteered in one turn.
+7. **Repair** is supported — corrections ("no, I said…") overwrite the targeted slot without losing surrounding context.
+8. The **`system-prompt` persona** (tone, scope, boundaries) is versioned and reviewed like production copy, and its voice is consistent across confirmations, errors, and refusals.
+9. **Refusal / limits messaging** is in-voice, specific, and redirects to an alternative rather than a blunt decline.
+10. The **chatbot opener states scope + capability** and offers suggested-reply chips; no blank-box-without-affordance entry state.
+11. **Voice-first onboarding** teaches invocation and capabilities by example ("you can say…") with consistent earcons for state.
+12. **Transcripts / captions** exist for voice, and chat is screen-reader-friendly (announced messages, focus management, dual audio/visual paths).

package/reference/registry.json

@@ -965,6 +965,13 @@
       "type": "heuristic",
       "phase": 36.1,
       "description": "Phase 36.1 Tier-3 domain pack — civic/government UI patterns: Section 508 + WCAG 2.1 AA hard floor, multi-language gov forms (EN/ES/zh-Hans), Plain Writing Act (grade 6-8), USWDS, session-timeout warnings, one-thing-per-page forms. Detection signals + Audit checklist for Step 0F + design-auditor."
+    },
+    {
+      "name": "conversational-ui",
+      "path": "reference/conversational-ui.md",
+      "type": "heuristic",
+      "phase": 36.3,
+      "description": "Phase 36.3 Tier-3 conversational-UI patterns (voice + chatbot): voice-flow no-input/no-match reprompts + confirmation + human handoff, multi-turn dialogue (context carryover, slot-filling, repair), prompt-as-UX (the assistant persona/tone/boundaries as a versioned design artifact), chatbot empty-states + suggested replies, voice-first onboarding, error recovery + accessibility (transcripts/captions). Carries Detection signals + an Audit checklist; loaded by design-context-builder for the conversational project type. CLI/REPL UX out of scope."
     }
   ]
 }