Commit graph

6 commits

Author SHA1 Message Date
Ajax Davis
47ddc4dea7 fix: refactor tool execution to use generateText with proper tool handling
Major changes:
- Fixed package executor URL protocol handling (add https:// if missing)
- Switched from streamText to generateText for proper tool execution
- AI now calls tool AND generates natural language response
- Tool results no longer show raw JSON metadata

Tool executor (tool-executor-agent.ts):
- Use generateText() instead of streamText() for tool execution
- Add system prompt to guide AI to summarize tool results
- Return result.text for natural language output
- Tool definition uses inputSchema (AI SDK v6 format)

Package executor:
- Add getSandboxUrl() to ensure URL has https:// protocol
- Fixes "Failed to parse URL" error in production

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-30 20:30:01 +10:00
Ajax Davis
78c72b85f3 feat: upgrade to AI SDK v6 beta and Zod v4 to fix tool schema errors
OpenAI was rejecting tool definitions with error "schema must be a JSON Schema of 'type: "object"'". This was caused by AI SDK v5 not properly converting Zod schemas to JSON Schema format.

**Changes:**
- Upgrade AI SDK from v5.0.104 to v6.0.0-beta.124
- Upgrade @ai-sdk/openai from v2.0.74 to v3.0.0-beta.22
- Upgrade Zod from v3.25.76 to v4.1.13 across all packages

**AI SDK v6 breaking changes:**
- Tool definition API: `parameters` renamed to `inputSchema`
- Removed `aiTool()` wrapper - use plain object with description, inputSchema, execute
- Streaming API: Use `textStream` async iterator instead of onChunk callback
- Zod schemas now properly converted to JSON Schema for OpenAI

**Zod v4 breaking changes:**
- `z.record()` now requires two arguments: `z.record(keySchema, valueSchema)`
- `z.enum()` params changed: `errorMap` removed, use `message` instead
- Type system improvements require explicit type parameters
- Fixed type errors in @tpmjs/env, @tpmjs/npm-client, @tpmjs/types

**Files changed:**
- apps/web/src/lib/ai-agent/tool-executor-agent.ts
  - Updated tool definition to use `inputSchema` instead of `parameters`
  - Removed `aiTool()` wrapper
  - Fixed streaming to use `textStream` iterator
- packages/env/src/index.ts
  - Updated type constraint from `z.ZodRawShape` to `Record<string, z.ZodTypeAny>`
- packages/npm-client/src/package.ts
  - Fixed `z.record()` calls to include both key and value schemas
  - Added type assertions for record indexing
- packages/types/src/tpmjs.ts
  - Changed `errorMap` to `message` in z.enum() calls

**Testing:**
-  Type-check passes
-  Production build succeeds
-  All routes compile correctly

This fixes the tool execution error where OpenAI rejected tool schemas with invalid format.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-30 19:14:01 +10:00
Ajax Davis
2ce37ae2aa feat: integrate Railway sandbox microservice for secure package execution
**Replace VM2 with Railway Microservice:**
- Remove VM2 dependency (incompatible with Next.js Turbopack bundling)
- Create Express sandbox service at `/services/sandbox-executor/`
- Uses isolated-vm for V8-level isolation with 128MB memory limit
- 10-second execution timeout with proper error handling

**Package Executor Client:**
- Rewrite `@tpmjs/package-executor` to call remote sandbox via HTTP
- Add `executePackage()`, `clearCache()`, `checkHealth()` functions
- Use AbortController for timeout handling
- Proper TypeScript type assertions for API responses
- Reads `SANDBOX_EXECUTOR_URL` from environment (defaults to localhost:3000)

**Sandbox Service Features:**
- `/execute` - Execute npm packages in isolated environment
- `/health` - Health check endpoint with service info
- `/cache/clear` - Clear npm package cache
- Package caching in `/tmp/.tpmjs-cache` for faster subsequent runs
- CORS support for web app integration
- Automatic ESM/CommonJS package detection

**Deployment Configuration:**
- Dockerfile with isolated-vm native dependencies (python3, make, g++)
- Railway.json with health checks and restart policies
- Environment variables: PORT, PACKAGE_CACHE_DIR, ALLOWED_ORIGINS
- Production URL: https://tpmjs-production.up.railway.app

**Integration:**
- Add SANDBOX_EXECUTOR_URL to .env.local
- Update Next.js config to mark package-executor as external
- Maintain existing API routes at `/api/tools/execute/[...slug]`

This architectural change enables secure package execution on Vercel
by moving sandboxing to a dedicated microservice on Railway.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-30 01:30:47 +10:00
Ajax Davis
d73db1c333 fix: remove VM2 sandboxing to resolve Next.js build errors
**VM2 Removal:**
- Remove VM2 dependency from package-executor
- Rewrite executor to use direct package execution with require()
- Add TODO comment for future sandboxing implementation

**Why this change:**
- VM2 requires runtime filesystem access to bridge.js which doesn't work with Next.js Turbopack bundling
- Even marking as serverExternalPackages fails because VM2 uses hardcoded file paths
- Direct execution allows builds to complete while we find Next.js-compatible sandboxing solution

**Next Steps:**
- Implement proper sandboxing with isolated-vm or similar Next.js-compatible solution
- Add security measures for package execution
- Consider moving package execution to separate microservice

This unblocks CI/CD while maintaining playground functionality.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-30 00:57:32 +10:00
Ajax Davis
ca57de533a fix: resolve Next.js routing and package-executor build errors
**API Route Restructuring:**
- Move execute endpoint from /api/tools/[...slug]/execute to /api/tools/execute/[...slug]
- Move simulations endpoint from /api/tools/[...slug]/simulations to /api/tools/simulations/[...slug]
- Fix Next.js App Router constraint: catch-all segments must be terminal
- Update ToolPlayground component to use new endpoint paths

**Package Executor Export Fix:**
- Remove .js extensions from exports in @tpmjs/package-executor
- Change from './types.js' to './types' for proper TypeScript resolution
- Change from './executor.js' to './executor' for proper TypeScript resolution
- Fixes "Export executePackage doesn't exist in target module" build error

**Next.js Configuration:**
- Add vm2 and @tpmjs/package-executor to serverExternalPackages
- Prevents bundling VM2 which requires filesystem access to internal files

**Code Quality:**
- Add biome-ignore for excessive complexity in SSE stream handling
- Add biome-ignore for decorative loading spinner SVGs (2 instances)

Note: VM2 sandboxing still has compatibility issues with Next.js Turbopack.
This may need to be replaced with a different sandboxing approach or disabled.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-30 00:52:56 +10:00
Ajax Davis
6eb51d1371 feat: add interactive tool playground with AI-powered execution
Implement comprehensive tool testing environment with real package execution, AI agents, and token tracking.

**New Features:**
- Interactive playground UI with 4 tabs (Input, Output, Logs, Token Usage)
- Real npm package execution in VM2 sandbox with security constraints
- AI-powered tool execution using AI SDK v5 and GPT-4 Turbo
- Server-Sent Events (SSE) streaming for real-time progress updates
- Comprehensive 4-category token tracking (Input, Tool Description, Schema, Output)
- Visual token breakdown with colored progress bars
- IP-based rate limiting (10 executions/hour per IP)
- Database persistence of all simulations with full metadata

**Database Schema:**
- New `Simulation` model for execution records
- New `TokenUsage` model for detailed token metrics
- New `ExecutionLog` model for execution event tracking
- Added simulations relation to Tool model

**Package Executor (@tpmjs/package-executor):**
- VM2 sandbox with 5-second timeout
- Blocked dangerous modules (fs, net, http, https, child_process)
- LRU file system cache in /tmp/.tpmjs-cache
- Package installation and caching strategy

**AI Agent Service:**
- TPMJS parameter to Zod schema conversion
- AI SDK tool definition generation
- Token counting using tiktoken library
- GPT-4 Turbo pricing estimation
- Streaming text execution with callbacks

**API Endpoints:**
- POST /api/tools/[...slug]/execute - SSE streaming execution
- GET /api/tools/[...slug]/simulations - Execution history
- Rate limit headers (X-RateLimit-Limit, X-RateLimit-Remaining)

**Frontend Components:**
- ToolPlayground - Main playground UI with tabs
- TokenBreakdown - Visual token metrics with colored bars
- Integrated above README section on tool detail pages

**Security:**
- VM2 sandboxing prevents filesystem/network access
- Rate limiting prevents abuse
- IP tracking for usage monitoring
- Timeout protection (60s max API duration, 5s VM timeout)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-30 00:36:05 +10:00