startCommand overrides Dockerfile CMD causing build failure. Removed
restartPolicyMaxRetries=-1 which is not a valid Railway value.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sync system was timing out because discovery endpoints (keyword, changes)
also ran schema extraction (~10-15s per tool). Now discovery is fast
(npm metadata + DB writes only) and a new /api/sync/enrich endpoint
handles schema extraction in time-budgeted chunks.
Railway executor was crashing without restarting due to unhandled promise
rejections, no restart policy, and no health checks. Added crash
protection, graceful shutdown, cache size limits, railway.toml with
ALWAYS restart policy, and upgraded Deno from 1.39 to 2.1.9.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extend schema sanitizer to handle two more invalid patterns found in
production: TypeScript-style type values like "string[]" and
"'markdown' | 'mdx'" (not valid JSON Schema), and properties with
conflicting type:"object" alongside oneOf containing string/array
types (e.g. resend sendEmail cc/to/bcc fields).
Tool schemas from the database can contain JSON Schema keywords that
Claude's API rejects (minimum, maxLength, top-level oneOf, old $schema
drafts, etc.), causing a 400 error that crashes the entire client
session. Add sanitizeInputSchema() that recursively strips unsupported
keywords before returning tools via MCP tools/list.
Tool schemas from the database can contain JSON Schema keywords that
Claude's API rejects (minimum, maxLength, top-level oneOf, old $schema
drafts, etc.), causing a 400 error that crashes the entire client
session. Add sanitizeInputSchema() that recursively strips unsupported
keywords before returning tools via MCP tools/list.
The -H/--header flag is variadic and swallows subsequent positional
args when placed before them. Move name and URL before flags so the
CLI parses correctly. Also update docs from npx mcp-remote to native
HTTP transport.
Switch collection MCP configs from npx mcp-remote to native HTTP
transport, fix Claude Code CLI arg order (options before name/url),
rename API key placeholder to YOUR_TPMJS_API_KEY, and add moltbook
social network tool to official blocks.
Restores Collections nav link in AppHeader, full collections/[id] detail
page with MCP URLs, dashboard collections Connect tab with McpUrlDisplay,
and CollectionDetailClient with ForkButton, CodeBlock, UseCases, and
Scenarios sections.
The /collections page was incorrectly replaced with a redirect to /
in a previous commit. This restores the full public collections
listing with search, sorting, virtualized table, likes, and copy.
- Add @tpmjs/tools-resend with email API tools and blocks.yml entries
- Add MCP route for collection skill discovery
- Add InstallationSection component for collections
- Refactor collection pages to use shared components and simplify layouts
- Update skills questions API, rate limiting, and API key handling
- Add tpmjs-tool-creator skill for Claude
- Update video feature scenes and fix lint issues
- Update .gitignore with IDE and temp file exclusions
Rebrand homepage messaging around toolspace virtualization concept:
- Hero: "INFINITE TOOLSPACE" with "A million tools. Zero configuration."
- New ToolspaceSection explaining the three-stage pipeline
- FeaturesSection emphasizes retrieval layer and selection at scale
- Integration section: "MCP is the socket, TPMJS is the OS"
- Publish section: "Join the Infinite Toolspace"
- Add Railway executor template with one-click deploy support
- Create Railway documentation page at /docs/executors/railway
- Update main executors page with Railway as official recommendation
- Add Railway to platform comparison table with new columns
- Update inter-page navigation for Railway → Unsandbox → Vercel flow
- Update FAQ to recommend Railway for most use cases
Railway executor features:
- Zero-dependency Node.js HTTP server
- Docker support via included Dockerfile
- Health check endpoint at /health
- Tool execution at /execute-tool
- API key authentication support
- Auto-restart on failure via railway.json
Restructure executor documentation to support multiple platforms:
- Main /docs/executors page now serves as overview with platform selector
- Add dedicated /docs/executors/unsandbox guide with CLI deployment
- Add dedicated /docs/executors/vercel guide with one-click deploy
- Include platform comparison table and shared API specification
- Add MCP Server Integration section with Claude Desktop, Cursor, VS Code config examples
- Add REST API Reference documenting /api/tools, /api/tools/search, /api/tools/execute endpoints
- Add "Building an Agent Like Omega" section with complete implementation examples
- Include SSE streaming patterns for real-time tool execution UI
- Update .gitignore to exclude .env*.local files
When a tool execution fails, return an error object instead of throwing.
This allows the AI model to see the error and inform the user properly,
rather than silently failing with no output.
- Changed @tpmjs/registry-execute and @tpmjs/registry-search to workspace:*
- This allows the web app to use local workspace versions
- Published npm packages still available for external users
- Updated all packages from ai@6.0.23 to ai@6.0.49
- Added pnpm override to ensure consistent version
- Created changeset for publishing affected packages
- @tpmjs/registry-search and @tpmjs/registry-execute use ESM syntax
- Move from serverExternalPackages to transpilePackages so Next.js
properly bundles and transpiles them
- Revert to static imports now that packages are bundled correctly
- Move @ai-sdk/devtools import to lazy dynamic import to prevent
module load failures in production environment
- Add try-catch around response.json() in client to handle
non-JSON error responses gracefully
- Display proper error message when server returns non-JSON response
- Light mode: pure white backgrounds (#FFFFFF) with high contrast
near-black text (#0A0A0A) for better readability
- Dark mode: clean neutral blacks (#0A0A0A base) instead of warm browns
- Borders use neutral grays instead of warm-tinted grays
- Status colors (success, warning, error, info) more saturated and vibrant
- Maintains copper accent color (#A6592D) as brand element
- Overall cleaner, more minimal feel with stronger contrast
- Create /api/skills/questions endpoint for listing questions with pagination
- Create /api/skills/questions/[id] endpoint for individual question details
- Add questions list page with filtering by skill
- Add question detail page with full answer, related tools, and similar questions
- Make activity feed cards clickable links to question detail
- Add "View all" link in SkillsSection
- Replace hardcoded purple/blue colors with primary/10 backgrounds
- Use text-primary for icons instead of hardcoded colors
- Use error tokens for error states instead of red-*
- Use ProgressBar component instead of custom div progress bars
- Use EmptyState component for empty activity feed
- Use Card components consistently with proper padding
- Fix TypeScript types in integration tests
Implements a skills endpoint that evolves through agent conversations:
- GET /:username/collections/:slug/skills - Returns markdown skill summary
- POST /:username/collections/:slug/skills - Ask questions, get RAG+LLM responses
Features:
- OpenAI text-embedding-3-large (3072 dims) for semantic similarity
- GPT-4.1-mini for response generation with RAG context
- Lazy seeding of synthetic questions on first access
- Cache hits for >95% similar questions
- Real-time skill graph updates (emergent skill taxonomy)
- Session support for multi-turn conversations
- Activity feed and stats APIs for UI
Database models: SkillQuestion, Skill, SkillSession, SkillQuestionSkill, SkillQuestionTool
Add abstract tool rendering infrastructure to @tpmjs/ui:
- ToolRenderer component with registry-based renderer lookup
- DefaultJsonRenderer as fallback with collapsible JSON display
- RegistrySearchRenderer for registrySearchTool results
- RegistryExecuteRenderer for registryExecuteTool results
- registerBuiltInRenderers() for idempotent initialization
Update Omega chat page to use new ToolRenderer:
- Replace inline ToolCallCard with ToolRenderer component
- Add helper functions to convert between ToolCall and ToolPart
- Remove unused expandedToolCalls state (managed internally)
Also add video file extensions to .gitignore
- Add /api/omega/settings/env-vars API endpoints for CRUD operations
- Add /omega/settings page with env var management UI
- Pass user env vars to tool execution in messages route
- Detect and warn about missing required env vars via SSE events
- Add EnvVarWarningBanner component to chat UI
- Update system prompt to document registry tools usage
- Improve auth flow on landing page with sign-in redirect
- Add Prisma models for conversations, messages, participants, tool runs, and user settings
- Create API endpoints for conversation CRUD and SSE message streaming
- Build landing page with sample prompts at /omega
- Build chat interface with real-time streaming at /omega/[conversationId]
- Integrate @tpmjs/registry-search and @tpmjs/registry-execute packages
- Use OpenAI GPT-4.1 Mini as the default model
- Create /:username/collections/:slug/usage.md endpoint
- Shows real-world usage patterns from test scenarios
- Filters out health check scenarios automatically
- Groups examples by tags for better organization
- Update skills.md to reference usage.md for usage examples
Implements batched generation to avoid Vercel 120s timeout for collections
with 60+ tools. Uses per-tool caching and recursive serverless invocations.
- Add ToolSkillsCache and SkillsGenerationJob models to schema
- Create tool-skills-generator.ts for per-tool markdown generation
- Create skills-summary-generator.ts for final pass summary/intro
- Update route handler with chunked generation logic
- Small collections (<20 tools) use original monolithic approach
- Large collections use 10-tool batches with progress tracking
