IdeaMesh

A multi-agent AI ideation platform that turns a raw software idea into a fully reasoned-out plan, with branchable pipelines and in-context refinement of any generated content.

Users feed in a seed idea (personal or business) and the platform runs a 6-stage (plus optional pitch deck) LLM pipeline: brainstorm → Q&A → SWOT + critique → competitor research → comparison → blueprint → pitch deck.

Every generated piece of content is editable in place: turn on Edit mode, click on a weakness, an audience, a build phase, a slide bullet, type feedback, queue it, then proceed — the platform refines all queued feedback in one coherent pass and (if the change would invalidate downstream stages) forks a new branch so the original timeline stays intact.

---

Getting started

npm install
npm run dev

Open http://localhost:3000.

Required environment variables

.env.local:

# LLM
LLM_API_KEY=sk-...
LLM_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-4o

# Web search is keyless: DuckDuckGo with Bing fallback.
SEARCH_PROVIDER=duckduckgo

# Supabase (auth + pipeline persistence)
NEXT_PUBLIC_SUPABASE_URL=https://<project>.supabase.co
NEXT_PUBLIC_SUPABASE_ANON_KEY=your-anon-key

The Supabase project must have the profiles and pipelines tables created (see context.md → Supabase schema). Google OAuth, if used, must be configured in the Supabase dashboard.

---

Headline features

1. The 6+1 ideation pipeline

#	Stage	Agents
1	Seed	User input (title, description, personal/business)
2	Mesh	Brainstorm, Q&A
3	Probe	Pros/Cons, Critique
4	Scout	Keyless DuckDuckGo search with Bing fallback
5	Compare	Competitor diff, gap analysis
6	Blueprint	Tech stack, build phases, MVP scope, coding-agent prompts
7	Pitch Deck	6–8 slide investor deck (business ideas only)

Each stage auto-runs on first visit, can be re-run manually, and writes its output back into the PipelineContext that downstream stages consume.

2. Branching (Copy-on-Write)

Every pipeline is a node in a tree:

• A pipeline can be forked at any stage. The fork deep-copies all upstream state, resets downstream stages to idle, and links to the parent via parentId. The root of the tree is tracked denormalised in rootId.
• Branches are visualised in the left sidebar as soft dashed limbs growing off the active spine.
• Click any branch chip to swap the active pipeline; the "Original timeline" chip jumps back to the root.
• Deleting a pipeline cascades to its descendants (Postgres FK ON DELETE CASCADE).

3. In-context refinement with queued notes

• Edit mode toggle in the header turns any wrapped block into a clickable target.
• Clicking opens a floating popover; the user types feedback and queues it. The block now wears a sticky-note chip showing the pending feedback.
• Multiple notes can be queued across an entire stage. A floating bottom bar shows the count and a Proceed with refinements button.
• On Proceed, the platform:
  1. forks the pipeline at the current stage (only if any downstream stage has already produced output — otherwise no fork),
  2. sends all queued notes to /api/agents/refine-batch in a single LLM call so refinements stay internally coherent,
  3. validates each refinement against its original JSON schema (no key add/remove, same primitive type, etc.),
  4. patches the context and saves once.

One batch refinement = at most one fork.

4. Mesh expansion opt-in

The Brainstorm agent emits N "expansions" (different angles for the idea). The user toggles each as in-scope or out-of-scope; only the selected ones are emphasised to downstream stages (PipelineContext.selectedExpansions). Editing Core Value or Target Audiences still goes through the standard refinement flow.

5. Auth + cloud persistence

Supabase Auth (email/password + Google OAuth) gates the workspace. Each user's pipelines are stored in public.pipelines with RLS isolating rows by user_id. Unauthenticated users get an ephemeral localStorage fallback for local prototyping.

---

Project layout

src/
  app/
    (workspace)/          # Route group: header + StageRail + edit/queue mountpoints
      mesh/   probe/   scout/   compare/   blueprint/   pitchdeck/
      layout.tsx
    api/agents/
      brainstorm/  qa/  probe/  scout/  compare/  blueprint/  pitchdeck/
      refine/             # Single-target refinement (legacy, currently unused by UI)
      refine-batch/       # Batched refinement called by the queue bar
    auth/                 # login / signup / forgot-password / reset-password / callback
    profile/              # Account + saved pipelines
    page.tsx              # Landing page
  components/
    pipeline/StageRail.tsx           # Tree-visualizer sidebar
    stages/{Mesh,Probe,Scout,...}Panel.tsx
    ui/AgentCard.tsx
    ui/EditableBlock.tsx             # Wraps editable regions; renders sticky notes
    ui/EditPopover.tsx               # Floating queue composer
    ui/RefinementQueueBar.tsx        # Bottom bar that drives the batch refine
    ui/UserMenu.tsx
  lib/
    ai/client.ts          # Lazy OpenAI client (getAI/getModel)
    ai/prompts.ts         # Shared context block + per-agent prompts
    jsonPath.ts           # parsePath / getAtPath / setAtPath for refinement
    supabase/{client,server}.ts
    utils.ts              # cn(), safeParseJSON(), formatDuration()
  store/
    pipelineStore.ts      # Zustand: pipeline + savedPipelines + branchPipeline
    editModeStore.ts      # Zustand: edit toggle + refinement queue
  types/pipeline.ts
  proxy.ts                # Next 16 proxy (auth session refresh + route protection)

---

Scripts

npm run dev      # next dev
npm run build    # next build (Turbopack)
npm run start    # next start
npm run lint     # eslint

Both tsc --noEmit and npm run lint must be clean before a change is considered done.

---

Architecture deep dive

See context.md for the full architecture, data model, persistence layer, prompt patterns, schema migrations, branching semantics, and gotchas.