Stun — Full Codebase Analysis

Date: March 12, 2026
Based on: IDEA.md + PRD.md + full codebase scan

TL;DR

The codebase is ~70% complete against the PRD. The core infrastructure, AI pipeline, and hybrid canvas are solid. The gaps are primarily around: Gemini Live API (never added), real-time collaboration (only polling), frontend node interactivity (nodes aren't editable), and several media/content node types (PDF, CSV, YouTube embeds).

1. What's Fully Implemented ✅

Infrastructure & Platform

Feature	Status	Location
Next.js 14 App Router (TypeScript)	✅ Working	`web/`
Express backend (TypeScript)	✅ Working	`backend/src/`
Firebase Authentication (JWT)	✅ Working	`backend/src/api/middleware/auth.middleware.ts`
Firestore board persistence	✅ Working	`backend/src/services/board.service.ts`
Google GenAI SDK (Gemini)	✅ Working	`backend/src/config/genai.ts`
Environment validation (Zod)	✅ Working	`backend/src/config/envVars.ts`
Helmet + CORS + Rate limiting	✅ Working	`backend/src/app.ts`, `ratelimit.middleware.ts`
Docker + Cloud Run deployment	✅ Files exist	`backend/Dockerfile`, `web/Dockerfile`, `infra/`
GCP Terraform infra	✅ Written	`infra/modules/`
Error handling middleware	✅ Working	`backend/src/api/middleware/error.middleware.ts`

Canvas System (Hybrid 3-Layer Architecture)

Feature	Status	Location
TLDraw as infinite workspace backbone	✅ Working	`web/components/canvas/TLDrawWorkspace.tsx`
Excalidraw as visual UI layer	✅ Working	`web/components/canvas/ExcalidrawLayer.tsx`
React Flow as knowledge graph engine	✅ Working	`web/components/canvas/ReactFlowGraphLayer.tsx`
Camera sync across all 3 layers	✅ Working	`web/lib/camera-sync.ts`
Canvas ↔ Excalidraw element mapping	✅ Working	`web/lib/canvas-mapping.ts`
CanvasRoot orchestrating all layers	✅ Working	`web/components/canvas/CanvasRoot.tsx`
Screenshot capture (html2canvas)	✅ Working	`web/hooks/useScreenshot.ts`
Image compression pipeline	✅ Working	`web/lib/image-compression.ts`

AI Planning Pipeline

Feature	Status	Location
POST `/ai/plan` endpoint	✅ Working	`backend/src/api/routes/ai.routes.ts`
Gemini multimodal reasoning (screenshot + nodes)	✅ Working	`backend/src/services/gemini.service.ts`
Structured JSON action plan output	✅ Working	`backend/src/validators/action.validator.ts`
Zod validation of action plan	✅ Working	`backend/src/validators/action.validator.ts`
Spatial context builder	✅ Working	`backend/src/services/context-builder.ts`
Intent parser (rule-based classifier)	✅ Working	`backend/src/services/intent-parser.ts`
Orchestrator service	✅ Working	`backend/src/services/orchestrator.service.ts`
Planner prompt (comprehensive, 9 action types)	✅ Working	`backend/src/prompts/planner.prompt.ts`
Position bounds clamping (zone-safe placement)	✅ Working	`backend/src/utils/spatial.ts`

Action Execution (Frontend)

Action Type	Status	Notes
`move`	✅ Working	Repositions React Flow node
`connect`	✅ Working	Creates edge between nodes
`highlight`	✅ Working	Visual border + shadow, auto-removes
`zoom`	✅ Working	(viewport needed)
`group`	✅ Working	Adds groupId + dashed border
`cluster`	✅ Working	Circular arrangement of nodes
`create`	✅ Working	Creates Excalidraw text element → RF node
`delete`	✅ Working	Removes node from graph
`transform`	✅ Working	Changes node type
`layout`	✅ Working	Calls `/layout/transform` API

Board CRUD & Access Control

Feature	Status	Location
Create board	✅ Working	`POST /boards`
List boards (owned + collaborated)	✅ Working	`GET /boards`
Get board	✅ Working	`GET /boards/:id`
Update board (autosave)	✅ Working	`PUT /boards/:id`
Delete board	✅ Working	`DELETE /boards/:id`
Board visibility (edit/view/private)	✅ Working	`PATCH /boards/:id/visibility`
Collaborator add/remove	✅ Working	`POST/DELETE /boards/:id/share`
Owner/edit/view/none access levels	✅ Working	`backend/src/services/boardAccess.service.ts`
Autosave (3s debounce)	✅ Working	`web/store/board.store.ts`

UI Components

Component	Status	Location
Landing page (EnterPage)	✅ Working	`web/components/EnterPage.tsx`
Boards list page	✅ Working	`web/app/boards/page.tsx`
Board page (`/board/[id]`)	✅ Working	`web/app/board/[id]/page.tsx`
Sign-in page (Google OAuth)	✅ Working	`web/app/signin/`
FloatingCommandBar (AI + voice + add node)	✅ Working	`web/components/ui/FloatingCommandBar.tsx`
SidePanel with AI quick-actions	✅ Working	`web/components/ui/SidePanel.tsx`
ShareDialog (visibility + collaborators)	✅ Working	`web/components/ui/ShareDialog.tsx`
MediaUploader (drag & drop + URL)	✅ Working	`web/components/ui/MediaUploader.tsx`
SearchBar	✅ Working	`web/components/ui/SearchBar.tsx`
PresenceIndicators	✅ Working	`web/components/ui/PresenceIndicators.tsx`
DeleteDialog	✅ Working	`web/components/ui/DeleteDialog.tsx`
LinkPreview	✅ Working	`web/components/ui/LinkPreview.tsx`
Toast notifications	✅ Working	`web/components/Toast.tsx`
ErrorBoundary	✅ Working	`web/components/ErrorBoundary.tsx`
ProfileIcon	✅ Working	`web/components/ProfileIcon.tsx`

Media System

Feature	Status	Location
File upload (GCS + local fallback)	✅ Working	`backend/src/services/media.service.ts`
Image thumbnail generation (sharp)	✅ Working	`media.service.ts`
Image/video/document node rendering	✅ Working	`web/components/nodes/MediaNode.tsx`
YouTube/Vimeo card (thumbnail + link)	✅ Working	`MediaNode.tsx`
URL parsing + link preview (cheerio)	✅ Working	`media.service.ts`
Drag & drop file upload	✅ Working	`web/hooks/useDragAndDrop.ts`
`POST /media/upload`	✅ Working	`backend/src/api/routes/media.routes.ts`
`POST /media/parse-url`	✅ Working	media routes
`GET /media/preview`	✅ Working	media routes

Search

Feature	Status	Location
Semantic search via `text-embedding-004`	✅ Working	`backend/src/services/search.service.ts`
Cosine similarity ranking	✅ Working	`search.service.ts`
`POST /search` endpoint	✅ Working	`backend/src/api/routes/search.routes.ts`
`useSearch` hook with 350ms debounce	✅ Working	`web/hooks/useSearch.ts`
Navigate-to-node on result select	✅ Working	`CanvasRoot.tsx` + `useSearch.ts`
Node highlight on search	✅ Working	`CanvasRoot.tsx`

Presence System

Feature	Status	Location
Heartbeat presence (15s interval)	✅ Working	`web/hooks/usePresence.ts`
Active users polling (10s interval)	✅ Working	`usePresence.ts`
`POST/GET /presence/:boardId`	✅ Working	`backend/src/api/routes/presence.routes.ts`
Stale presence cleanup	✅ Working	`backend/src/services/presence.service.ts`
Presence indicator UI	✅ Working	`PresenceIndicators.tsx`

Layout Transformations

Layout Type	Status	Location
Mind map (radial)	✅ Working	`backend/src/services/layout.service.ts`
Roadmap (sequential phases)	✅ Working	`layout.service.ts`
Timeline (chronological)	✅ Working	`layout.service.ts`
Flowchart (hierarchical)	✅ Working	`layout.service.ts`
Presentation (slide-by-slide)	✅ Working	`layout.service.ts`
`POST /layout/transform`	✅ Working	`backend/src/api/routes/layout.routes.ts`

Voice Input

Feature	Status	Location
Web Speech API STT	✅ Working	`web/hooks/useVoice.ts`
Mic button → transcript → command input	✅ Working	`FloatingCommandBar.tsx`
Continuous + interim results	✅ Working	`useVoice.ts`

2. Partially Implemented / Has Bugs ⚠️

2.1 Gemini API vs Vertex AI — Critical Hackathon Issue

Status: ⚠️ Using Gemini API directly, NOT Vertex AI SDK

backend/src/config/genai.ts uses GoogleGenAI({ apiKey: envVars.GEMINI_API_KEY })
This is the Gemini API (ai.google.dev), not Vertex AI (cloud.google.com/vertex-ai)
PRD compliance checklist says: ✔ Uses Vertex AI — but the code doesn't
The VERTEX_MODEL env var name is misleading — it just sets the model name, not actual Vertex AI
Impact: Hackathon "Uses Vertex AI" requirement may not be satisfied
Fix needed: Use VertexAI from @google-cloud/vertexai OR googleai provider on Vertex AI endpoint

2.2 React Flow Nodes Not Interactive

Status: ⚠️ Nodes are read-only

ReactFlowGraphLayer.tsx sets nodesDraggable={false}, nodesConnectable={false}, elementsSelectable={false}
The layer's pointerEvents: "none" means ALL user interactions pass through to Excalidraw
TextNode has no inline editing — just displays data.label statically
Users cannot drag React Flow nodes manually — only AI can move them
Impact: Users can't manually rearrange AI-created nodes on the graph layer

2.3 Voice Input Is One-Way (No AI Voice Response)

Status: ⚠️ STT only, no TTS

useVoice.ts captures voice → sets transcript in input field
User still must press Enter or "Send" to trigger AI
No AI voice response back (no text-to-speech)
IDEA/PRD says: "Speaks optionally (voice)" — NOT implemented
Impact: Voice is a convenience input mechanism, not true voice-first interaction

2.4 YouTube/Video Nodes — No Actual Embedding

Status: ⚠️ Shows thumbnail + button, opens in new tab

MediaNode.tsx renders a YouTube/Vimeo card with thumbnail
Clicking opens the video in a new browser tab
No in-canvas YouTube <iframe> embed
No video keyframe extraction
No transcript summarization from video content
Impact: Video content can't be analyzed or played within the canvas

2.5 Screenshot in usePlanAndExecute — Partial

Status: ⚠️ Screenshots of hybrid canvas are unreliable

usePlanAndExecute.ts uses html2canvas on #canvas-root
TLDraw and Excalidraw render on Canvas elements — html2canvas often captures blank/empty areas
The screenshot passed to Gemini may not accurately represent what the user sees
Impact: AI's multimodal reasoning may be based on an empty/incorrect screenshot

2.6 Layout Action in Backend Validator — Missing Schema

Status: ⚠️ layout action type not in Zod schema

action.validator.ts defines union for: move, connect, highlight, zoom, group, create, delete, transform
layout action type is in the prompt and frontend executor but NOT in the backend Zod schema
If AI returns a layout action, backend validation will reject it with a Zod error
Fix needed: Add layoutActionSchema to actionSchema union in validator

2.7 Cluster Action Not in Backend Validator

Status: ⚠️ Same issue as layout — cluster is in executor but not in Zod schema

Backend will reject cluster action type
Only the frontend executor has it

2.8 TextNode Label — Not Editable in UI

Status: ⚠️ Static render only

TextNode renders data.label with no edit affordance
No double-click to edit, no contenteditable
Nodes created by AI via create action show static text
Impact: Users can't correct or update AI-created node content

2.9 CanvasRoot Not Wired to MediaUploader Properly

Status: ⚠️ Upload works but nodes aren't added to canvas

MediaUploader calls onMediaUploaded callback
In CanvasRoot.tsx the showMediaUploader state is managed but the onMediaUploaded callback implementation needs verification for actually creating Media nodes on the canvas

3. Missing — Not Implemented ❌

3.1 Gemini Live API (Real-Time Voice + AI)

Status: ❌ Not implemented

The IDEA.md explicitly mentions: "Gemini Live API (voice + real-time)"
Currently: Web Speech API → typed command → HTTP request → AI response
Should be: Continuous voice stream → Gemini Live bidirectional streaming → real-time AI narration + canvas actions
Impact: The product's core "voice-first spatial AI" interaction is a basic STT input, not live AI voice

3.2 Real-Time Collaboration (WebSockets / Firestore Listeners)

Status: ❌ Not implemented — only polling

Presence is polling every 10-15s
No WebSocket connection for live canvas state sync
If User A moves a node, User B sees it only after ~10+ seconds (if at all)
Impact: "Real-time collaboration" claim is misleading — it's async polling only

3.3 AI Auto-Organization (Background Continuous Scanning)

Status: ❌ Not implemented

IDEA.md says: "AI continuously scans all nodes and clusters related nodes, draws connections..."
Currently: AI only acts when user explicitly sends a command
No background AI worker, no automatic clustering trigger
Impact: Canvas doesn't self-organize, which was a core IDEA feature

3.4 OCR for Screenshots / Image Content Analysis

Status: ❌ Not implemented

IDEA.md says: "OCR for screenshots", "Image scanning & annotation"
Currently: Images are uploaded and stored, but no AI analysis of image content
No endpoint to analyze/annotate an existing image node
Impact: Images on canvas are opaque to the AI (can't read text in images)

3.5 PDF Node / PDF Viewer

Status: ❌ Not implemented

IDEA says: Drop PDFs → becomes a Node with visual preview
Backend media service handles PDF upload but no PDF rendering
No PDF viewer component
No PDF text extraction for AI context

3.6 CSV / Sheets Node

Status: ❌ Not implemented

IDEA says: Drop CSV/Sheets → becomes a node
Media uploader accepts CSV files but no table/spreadsheet node renderer
No CSV parsing or visualization component

3.7 TopBar Navigation (Figma-Style)

Status: ❌ Not implemented

IDEA.md Section 5.7: "Asset Navigation Bar (Figma-like): Media asset library, Search, Voice input, AI tools, Canvas tools (draw, connect, resize)"
PRD structure shows components/layout/TopBar.tsx — does not exist
Currently only FloatingCommandBar (bottom) + SidePanel (right)
No persistent top navigation bar

3.8 Node Right-Click Context Menu

Status: ❌ Not implemented

No right-click context menu on React Flow nodes
Users can't delete, edit, duplicate, or AI-analyze specific nodes via context menu
All node operations must go through the command bar or AI

3.9 Video Keyframe Extraction / Transcript Summarization

Status: ❌ Not implemented

IDEA.md says: "Video keyframe extraction", "Transcript summarization"
MediaNode just shows YouTube thumbnail — no actual video processing
No backend pipeline for video content analysis

3.10 Semantic Embedding Storage / Persistent Search Index

Status: ❌ Not fully implemented

Search service generates embeddings on-the-fly for every search query
No persistent vector store or cached embeddings
Every search re-embeds all nodes from scratch
Impact: Search is slow for large boards; embedding costs accumulate

3.11 Node History / Undo System

Status: ❌ Not implemented

No undo/redo for AI actions
If AI deletes wrong nodes, no recovery
Autosave overwrites previous state
Impact: High risk of destructive AI actions with no recovery

3.12 AI Memory / Canvas Memory

Status: ❌ Not implemented

IDEA.md Section 5.7 SidePanel: "History, Memory" listed as UI sections
No persistent AI memory between sessions
AI has no context of what it did before in the current session
Impact: AI can't reference previous actions or maintain context across commands

3.13 Diagram Generation Node Type

Status: ❌ Partially missing

Planner prompt describes diagram creation with start/process/decision/end nodes
But transform action's "diagram" nodeType has no corresponding DiagramNode component
No Mermaid diagram rendering in React Flow nodes
Impact: transform to "diagram" produces no visual change

3.14 Node Annotations (Bounding Box Overlays)

Status: ❌ Not implemented

IDEA.md says: "Draws bounding boxes on images", "Adds annotation nodes"
No annotation overlay system on image nodes
No bounding box drawing from AI actions

3.15 Continuous Live Observation Mode

Status: ❌ Not implemented

PRD Future Roadmap: "Continuous live observation mode"
AI can't passively watch the canvas and proactively suggest actions

3.16 Mobile Responsiveness

Status: ❌ Out of scope per PRD, but worth noting

4. Working Status by System

System                          Status    Confidence
─────────────────────────────────────────────────────
Authentication (Firebase)        ✅ WORKS    HIGH
Board CRUD (Firestore)           ✅ WORKS    HIGH
AI planning pipeline (Gemini)   ✅ WORKS    HIGH
Canvas rendering (3-layer)       ✅ WORKS    HIGH
AI action execution              ✅ WORKS    MEDIUM
Media upload (local mode)        ✅ WORKS    HIGH
Media upload (GCS mode)          ✅ WORKS    MEDIUM
Semantic search                  ✅ WORKS    HIGH
Layout transformations           ✅ WORKS    HIGH
Voice STT input                  ✅ WORKS    HIGH
Presence/heartbeat               ✅ WORKS    HIGH
Collaboration (share)            ✅ WORKS    HIGH
─────────────────────────────────────────────────────
Vertex AI compliance             ⚠️  ISSUE   HIGH RISK
Backend action validation        ⚠️  ISSUE   MEDIUM RISK
Screenshot for AI                ⚠️  PARTIAL MEDIUM RISK
Node interactivity               ⚠️  PARTIAL MEDIUM RISK
YouTube embedded player          ⚠️  PARTIAL LOW RISK
─────────────────────────────────────────────────────
Gemini Live API (voice stream)   ❌ MISSING  HIGH PRIORITY
Real-time collab (WebSocket)     ❌ MISSING  MEDIUM
PDF viewer node                  ❌ MISSING  MEDIUM
CSV/table node                   ❌ MISSING  LOW
AI auto-organization             ❌ MISSING  MEDIUM
OCR / image analysis             ❌ MISSING  MEDIUM
Node undo/redo                   ❌ MISSING  MEDIUM
TopBar navigation                ❌ MISSING  LOW
AI memory                        ❌ MISSING  LOW
Video transcript                 ❌ MISSING  LOW

5. Critical Bugs to Fix Now

Bug 1: Missing `layout` and `cluster` in backend action validator

File: backend/src/validators/action.validator.ts
Problem: AI can return layout or cluster action types (they're in the prompt), but backend Zod schema rejects them → 400 error
Fix: Add these two schemas to the actionSchema discriminated union

const clusterActionSchema = z.object({
  type: z.literal("cluster"),
  nodeIds: z.array(z.string()).min(1),
  center: z.object({ x: z.number(), y: z.number() }).optional(),
});

const layoutActionSchema = z.object({
  type: z.literal("layout"),
  layoutType: z.enum(["mindmap", "roadmap", "timeline", "flowchart", "presentation"]),
  options: z.object({
    spacing: z.object({ x: z.number(), y: z.number() }).optional(),
    centerPosition: z.object({ x: z.number(), y: z.number() }).optional(),
    direction: z.enum(["horizontal", "vertical", "radial"]).optional(),
  }).optional(),
});

Bug 2: Gemini API vs Vertex AI

File: backend/src/config/genai.ts
Problem: Using Gemini API key, not Vertex AI — hackathon compliance violation
Fix: Switch to @google-cloud/vertexai SDK or use the google-genai SDK's Vertex AI constructor

Bug 3: Screenshot reliability

File: web/hooks/usePlanAndExecute.ts
Problem: html2canvas on #canvas-root doesn't capture TLDraw/Excalidraw canvas frames correctly (they're rendered as WebGL/Canvas2D contexts)
Fix: Use getScreenshot() API from Excalidraw instance, OR use TLDraw's editor.toImage(), store it in a ref, pass to AI planner

Bug 4: `reasoning` field missing from action plan schema

File: backend/src/validators/action.validator.ts
Problem: Frontend usePlanAndExecute reads plan.reasoning for toast message, but actionPlanSchema doesn't include reasoning field — Zod strips unknown fields by default
Fix: Add reasoning: z.string().optional() to actionPlanSchema

6. What Needs to Be Added for Full PRD Compliance

Priority 1 — Hackathon Blockers

Fix Vertex AI usage — Switch to proper Vertex AI SDK (requirement in PRD compliance checklist)
Fix backend validator — Add layout and cluster action schemas
Fix reasoning field — Pass through from AI for UX toasts
Validate screenshot — Use Excalidraw API or fallback blank image reliably

Priority 2 — Core UX Missing

Inline node editing — Double-click to edit TextNode label
Node right-click context menu — Delete, duplicate, AI-analyze
Undo/redo system — Before/after snapshots for AI action reversibility
TopBar — Media asset panel, canvas tools (currently missing entirely)

Priority 3 — PRD Must-Haves not done

Gemini Live API — Real-time voice bidirectional stream (huge scope, likely post-hackathon)
YouTube iframe embed — Proper embeds instead of thumbnail + open-in-tab
AI response voice (TTS) — Text-to-speech for AI reasoning output

Priority 4 — PRD Nice-to-Haves

Persistent embeddings — Cache node embeddings in Firestore for search speed
AI auto-organization trigger — Manual "auto-organize" button calling the AI planner
PDF text extraction — Backend pdf-parse for AI context
CSV table node — Basic table renderer

7. Architecture Quality Assessment

Strengths

The 3-layer hybrid canvas (TLDraw + Excalidraw + React Flow) is architecturally sophisticated and well-documented
The AI pipeline (intent parser → spatial context → Gemini → Zod validation → action executor) is clean and extensible
Backend is properly layered (controllers → services → validators)
Security is strong: JWT auth, Zod validation, coordinate clamping, node ID whitelist
Autosave with debounce is well-implemented

Weaknesses

The 3-layer canvas creates complexity: camera sync is fragile across 3 different viewport systems
React Flow layer is purely infrastructure (all nodes invisible, no pointer events) — feels like over-engineering for current use
Voice is an afterthought: it just fills a text box, no actual AI voice integration
No undo stack means every AI action is irreversible — high risk in a demo
html2canvas screenshots are unreliable for canvas-based rendering

8. File/Component Count Summary

Area	Files
Backend services	10
Backend controllers	8
Backend routes	9
Backend middleware	4
Frontend components	~20
Frontend hooks	10
Frontend lib utilities	8
Frontend store	3
Tests (backend)	7
Infra (Terraform)	6+ modules

Total codebase files: ~100+ TypeScript files
Estimated LOC: ~8,000–12,000 lines

FilesExpand file tree

CODEBASE_ANALYSIS.md

Latest commit

History

CODEBASE_ANALYSIS.md

File metadata and controls

Stun — Full Codebase Analysis

TL;DR

1. What's Fully Implemented ✅

Infrastructure & Platform

Canvas System (Hybrid 3-Layer Architecture)

AI Planning Pipeline

Action Execution (Frontend)

Board CRUD & Access Control

UI Components

Media System

Search

Presence System

Layout Transformations

Voice Input

2. Partially Implemented / Has Bugs ⚠️

2.1 Gemini API vs Vertex AI — Critical Hackathon Issue

2.2 React Flow Nodes Not Interactive

2.3 Voice Input Is One-Way (No AI Voice Response)

2.4 YouTube/Video Nodes — No Actual Embedding

2.5 Screenshot in usePlanAndExecute — Partial

2.6 Layout Action in Backend Validator — Missing Schema

2.7 Cluster Action Not in Backend Validator

2.8 TextNode Label — Not Editable in UI

2.9 CanvasRoot Not Wired to MediaUploader Properly

3. Missing — Not Implemented ❌

3.1 Gemini Live API (Real-Time Voice + AI)

3.2 Real-Time Collaboration (WebSockets / Firestore Listeners)

3.3 AI Auto-Organization (Background Continuous Scanning)

3.4 OCR for Screenshots / Image Content Analysis

3.5 PDF Node / PDF Viewer

3.6 CSV / Sheets Node

3.7 TopBar Navigation (Figma-Style)

3.8 Node Right-Click Context Menu

3.9 Video Keyframe Extraction / Transcript Summarization

3.10 Semantic Embedding Storage / Persistent Search Index

3.11 Node History / Undo System

3.12 AI Memory / Canvas Memory

3.13 Diagram Generation Node Type

3.14 Node Annotations (Bounding Box Overlays)

3.15 Continuous Live Observation Mode

3.16 Mobile Responsiveness

4. Working Status by System

5. Critical Bugs to Fix Now

Bug 1: Missing layout and cluster in backend action validator

Bug 2: Gemini API vs Vertex AI

Bug 3: Screenshot reliability

Bug 4: reasoning field missing from action plan schema

6. What Needs to Be Added for Full PRD Compliance

Priority 1 — Hackathon Blockers

Priority 2 — Core UX Missing

Priority 3 — PRD Must-Haves not done

Priority 4 — PRD Nice-to-Haves

7. Architecture Quality Assessment

Strengths

Weaknesses

8. File/Component Count Summary

Bug 1: Missing `layout` and `cluster` in backend action validator

Bug 4: `reasoning` field missing from action plan schema