THE BREAKTHROUGH: AI SDK v6 tools use jsonSchema() which wraps plain JSON Schema objects, NOT Zod schemas. JSON Schema is fully serializable.
Changes:
1. Railway server: Extract raw JSON Schema from toolModule.inputSchema?.schema
2. Playground loader: Wrap received JSON Schema with { type: 'json_schema', schema: ... }
3. This matches AI SDK v6 format exactly - no Zod serialization needed
How it works:
- Tools define inputSchema: jsonSchema({ type: 'object', properties: {...} })
- AI SDK stores it as { type: 'json_schema', schema: {...} }
- Railway extracts the plain JSON Schema (.schema property)
- Sends it as plain JSON (fully serializable)
- Playground wraps it back in AI SDK format
- OpenAI receives valid JSON Schema for function calling
This fixes both errors:
✅ No more "def.shape is not a function" (not using Zod)
✅ No more "Invalid schema type None" (proper JSON Schema provided)
Note: Tools using Zod instead of jsonSchema() will need to migrate.
TPMJS standard: All tools MUST use jsonSchema() with plain JSON Schema.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Implements a complete dynamic tool loading system that allows the playground to discover and load tools from the TPMJS registry at runtime.
**Architecture:**
- Search tool package (@tpmjs/search-registry) - Searches registry for tools
- Search API endpoint (/api/tools/search) - Text-based search with scoring
- Pre-flight tool loading - Automatically searches and loads tools on every message
- Railway executor service (Deno) - Loads tools from esm.sh via HTTP imports
- Dynamic tool loader - Calls Railway to load and execute tools remotely
**Key Components:**
1. Railway Executor (apps/railway-executor/)
- Deno-based service that natively supports HTTP imports
- Endpoints: /load-and-describe, /execute-tool, /cache/stats, /cache/clear
- Deploys to Railway with deno run --allow-net --allow-env server.ts
2. Search Tool Package (packages/tools/search-registry/)
- AI SDK v6 tool for searching TPMJS registry
- Uses jsonSchema + inputSchema pattern
- Searches /api/tools/search endpoint
3. Search API (apps/web/src/app/api/tools/search/)
- Text-based search with composite scoring
- Scores: text relevance + quality boost + download boost
- Returns tool metadata with importUrl for dynamic loading
4. Dynamic Tool Loader (apps/playground/src/lib/dynamic-tool-loader.ts)
- Calls Railway service to load tools from esm.sh
- Creates tool wrappers that execute remotely
- Process-level module cache + per-conversation tracking
5. Pre-flight Loading (apps/playground/src/app/api/chat/route.ts)
- Automatically searches for tools on every user message
- Loads top 5 matching tools before agent processes request
- Merges with static tools for seamless experience
**Technical Decisions:**
- Deno over Node.js: Native HTTP import support without flags
- Remote execution: Tools run in Railway sandbox, not Vercel
- Pre-flight loading: Better UX than two-turn search pattern
- Text search: BM25 had dependency issues, simple scoring works well
**Environment Variables:**
- RAILWAY_SERVICE_URL: https://endearing-commitment-production.up.railway.app🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
