Skip to main content
ThinkEx is a Cloudflare Worker app managed by Vite+. The repo pins Node 24 and expects the Vite+ managed environment for local work.

​
Prerequisites

  • Node.js 24 (>=24.11.0 <25) through Vite+ managed mode
  • pnpm 11
  • Docker, only for the full dev profile (see Dev server profiles)
  • Infisical access for core team members, or a local .dev.vars file for contributors
If node, pnpm, or Corepack resolve to unexpected versions, run:

​
Install

​
Run with Infisical

Core team members with Infisical access can run:
This injects dev:/app secrets and starts the dev server.

​
Run without Infisical

Contributors can use a local .dev.vars file:
Only two variables are required to boot the app:
The auth page shows Continue as guest in local development, so Google OAuth credentials are not required for normal UI testing.

​
Dev server profiles

pnpm dev and pnpm serve:dev run a lite profile by default:
  • Every Cloudflare binding is emulated locally. Nothing is proxied to the real account, so a normal session incurs no Workers AI, Browser Rendering, R2, or Email charges — and no invite email is delivered to a real inbox.
  • Containers are off, so Docker is never invoked and the three container images are neither built nor started.
  • Miniflare’s local trace store is off, which is otherwise the largest thing in .wrangler/.
Remote-only features are unavailable in this profile: browser rendering, PDF export, image-to-text extraction, and email invite delivery. Core workspace, document, and auth flows all run locally.
The lite profile is not “free” — it removes the charges you did not ask for, not the ones you trigger.Only Cloudflare bindings are forced local. Three paid services are reached over plain HTTPS with an API key rather than through a binding, so no dev profile affects them:
  • AI chat, thread titles, and compaction — Vercel AI Gateway, billed per turn whenever AI_GATEWAY_API_KEY is set
  • web_search and research tools — Firecrawl
  • PDF text extraction — LlamaParse
Under pnpm dev, Infisical supplies those keys, so sending a chat message costs money in either profile. For a session that spends nothing at all, use pnpm serve:dev with a .dev.vars that omits those keys — the AI gateway then raises a clear error instead of billing.
When you need the production-shaped environment — testing image extraction, browser rendering, real R2 behavior, or anything running in a container — opt in explicitly:
The full profile bills real usage against the Cloudflare account, and sends real email. THINKEX_DEV_PROFILE=full is equivalent to --full for non-interactive callers.

​
Database

Local development requires PostgreSQL. Both development commands create an isolated database from CONDUCTOR_PORT and CONDUCTOR_WORKSPACE_ID (or port 3000 outside Conductor), apply pending migrations, and pass its direct URL to Wrangler’s local Hyperdrive binding automatically. Including the stable workspace ID prevents a newly created workspace from inheriting data or migration history when Conductor reuses a port:
Local Wrangler connects directly to PostgreSQL; it does not emulate Hyperdrive pooling or caching. The generated connection string is per-workspace local configuration and should not be stored in Infisical. If CLOUDFLARE_HYPERDRIVE_LOCAL_CONNECTION_STRING_HYPERDRIVE is already set, the development command uses it and deliberately skips automatic creation and migrations.

​
Docker

Docker is only needed for the full profile (pnpm dev:full), which declares Cloudflare Containers for code execution and file conversion — the first such startup may build large images. The default lite profile skips containers entirely, so Docker does not need to be running.

​
Reclaim local disk

Local development artifacts accumulate per checkout and are never pruned automatically — Miniflare’s trace store, locally built container images, and one PostgreSQL database per port and workspace ID ever used. To see what is reclaimable:
That is a dry run. Add --yes to apply it, and optionally:
  • --databases — also drop idle thinkex_* databases and their roles. Never drops this checkout’s database, and never one with an open connection, so a running worktree is safe.
  • --docker-cache — also prune Docker’s build cache. This is machine-wide, so other projects lose their cached layers too.
Your local application data is never touched: .wrangler/state/v3/do (Durable Object storage — workspaces, documents, AI threads) and .wrangler/state/v3/r2 (uploaded files) exist nowhere else, so no flag removes them.

​
Validate Changes

This runs Vite+ check, test, and build tasks through the task cache, which fingerprints inputs and replays previous results instead of redoing work. Measured on an M-series laptop: So pnpm verify is the right habit for “is everything still OK”, especially across several worktrees running the same checks on mostly-unchanged code. It is slower than the raw runners on a full cache miss, because it also fingerprints inputs and archives outputs — so while actively iterating, reach for the narrower commands instead:
The workers test project is the slow part and is dominated by module transform and import inside workerd, one runtime per test file. It already runs in parallel across cores; raising --maxWorkers past the default makes it slower, not faster. pnpm doctor runs React Doctor against changed files. Use pnpm doctor:full when you intentionally want a full React scan.