v1.13.0

v1.13.0 — Custom URL Routing & System Logging

🚀 New Features

SiteRoute — Custom URL Routing

• SiteRoute model — new database entity for SEO-friendly custom paths
  • path — unique full URL path (e.g. /my-article)
  • slug — short segment identifier
  • parentId — hierarchical structure support
  • kBConceptId — 1:1 link to KBConcept
• Next.js fallback routing — arbitrary paths handled via /_fallback/[...path].tsx
• Virtual uri field — KBConcept returns custom path or default /concepts/{id}
• Link components — ConceptLink, PostLink, TaskLink, WorkLogLink use uri for SEO-friendly URLs
• GraphQL API — siteRoute, siteRoutes, createSiteRoute resolvers

SystemLog

• SystemLog model — centralized logging for system events and errors

Task Management

• Create and edit tasks — full CRUD support for tasks
• Create task link — quick access to task creation

Chat Enhancements

• User current page data — chat messages now include user's current page context
• Autoclose and autoexpanded handlers — improved chat modal behavior
• useStopPropagationScroll hook — prevent scroll event bubbling

Image Generation

• More models — added additional image generation models
• ImageGenerator component update — improved UI and functionality

Markdown

• FilesUploader — markdown editor now supports file uploads

🔧 Technical Changes

Infrastructure

• Self-signed HTTPS for localhost — Traefik websecure entrypoint with TLS support
• Mailhog dev port — exposed port 1025 for local email testing

GraphQL Codegen

• TypedDocumentNode — automatic type inference in Apollo hooks without explicit generics
• Simplified syntax — components no longer need explicit type parameters

SEO

• LocalBusiness JsonLd — added structured data schema
• JsonLd breadcrumbs fix — corrected breadcrumb generation

UI/UX

• dvh units — replaced vh with dvh for better mobile viewport handling
• Auth form fix — resolved authentication form issues
• Authed redirect fix — corrected redirect behavior for authenticated users

Code Quality

• findUnique over findFirst — improved query precision
• PrismaContext req types — fixed request type definitions
• Console.log cleanup — removed debug statements

🐛 Bug Fixes

• Fix 404 errors handling
• Fix skills loading
• Fix markdown links
• Fix worklogs link
• Fix tasks searchable
• Fix edit task status
• Fix site origin detection
• Fix AppInitialProps types
• Update tasks page fetch policy

📦 Dependencies

• Upgraded project dependencies

---

Full Changelog: v1.12.0...v1.13.0

v1.12.0 — Infrastructure & Image Support

🚀 New Features

Post & KBConcept Images

• Post::image — posts now support an optional image field for visual content.
• KBConcept::image — knowledge base concepts can have associated images.
• KBConceptFiles relations — fixed relationships between concepts and files.

Image Generator

• ImageGenerator — new utility for AI-powered image generation.
• LLM client refactoring — improved architecture for LLM integrations with image generation support.

Mermaid Diagrams

• Mermaid support — added Mermaid diagram rendering capability for documentation and content.

Email System

• sendMail resolver — new GraphQL mutation for sending emails via the mail server.
• Mailserver deployment fix — resolved deployment issues with the mail server configuration.

Agent Enhancements

• Skills info in Agent Data — agent data now includes skills information for better context.
• readWebPage utility — new Utils/readWebPage helper for fetching and parsing web content.

🔧 Technical Changes

Infrastructure

• Node 22 migration — migrated base image to node:22-bookworm for improved performance and security.
• Varnish cache — added Varnish cache layer for static assets to improve load times.
• Traefik logs fix — resolved logging issues in Traefik configuration.
• Crypto config — moved cryptographic configuration to environment variables for better security.

Database

• File::CreatedBy rename — renamed the CreatedBy field in the File model for consistency.

UI Fixes

• Chat width — fixed chat widget width issues.
• ChatWidget scroll — resolved scrolling behavior in the chat widget.
• :disabled styles — fixed styling for disabled form elements.

Code Quality

• hasFieldInSelection helper — added utility for checking field presence in GraphQL selections.
• Translation to English — translated various components to English.

📦 Dependencies

No major dependency changes in this release.

---

Full Changelog: v1.11.0...v1.12.0

v1.11.0 — Skills Resolver & Browser Compatibility

🚀 New Features

File-Based Skills Resolver

A new GraphQL module that exposes a catalog of agent skills backed by files
on disk. Skills are designed to be stored next to the GraphQL types they
operate on and to ship with optional auxiliary files (markdown templates,
shell scripts, etc.).

• Discovery — the registry recursively scans server/schema/types/ for
  files literally named skillManifest.ts. Skills can therefore live under
  any graph-type folder, not only under Skills/skills/.
• Identifier — the skill id is derived from the manifest folder path
  relative to server/schema/types/ (POSIX form). Authors never assign ids
  manually.
• Manifest contract — each skillManifest.ts must export a named const
  skillManifest: SkillManifest. Invalid or missing manifests are skipped
  (with a console.error in development) so a half-written file cannot
  break the registry.
• Manifest fields — name, flat description, optional
  buildContent(ctx) for lazy markdown content (with full PrismaContext
  access), an optional auxiliary file list, and an optional executable
  (shell or node) with a typed argsSchema.
• Lazy content — Skill.content is resolved only if the client selects
  it, so listing skills stays cheap.
• Executor — child_process.spawn with a 60s timeout; declared args are
  forwarded as --key value pairs. Host-side execution is intentional for
  now; an isolated Docker runner is planned (task 001--docker-service).
• Helpers — renderTemplate({{path.to.value}}) for simple runtime
  substitution. A full MDX runtime is intentionally not wired up yet.
• Example — ships a hello-world skill demonstrating manifest, markdown
  template using DB user count, and a run.sh.

GraphQL surface:

type Skill {
  id: ID!
  name: String!
  description: String!
  content: String          # lazy — calls buildContent only if requested
  files: [String!]!
  hasExecutable: Boolean!
}

type Query {
  skills: [Skill!]!
  skill(id: ID!): Skill
}

type Mutation {
  executeSkill(id: ID!, args: Json): SkillExecutionResult!
}

See wiki/skills for the full guide.

Browser Compatibility & E2E Testing

End-to-end coverage for older browsers and a stricter bundle baseline.

• Browserslist — production target now covers Safari/iOS 13+ and modern
  Chrome/Firefox/Edge, so Next.js/SWC transpiles the client bundle for
  older runtimes.
• check:bundle — new npm script using es-check to verify the
  production bundle syntax against the supported baseline (ES2019).
• Playwright + WebKit — desktop Safari and mobile iPhone 13 projects
  catch runtime errors that es-check cannot detect (missing APIs,
  hydration failures, broken deps). baseURL is read from
  PLAYWRIGHT_BASE_URL (defaults to http://localhost:3000); the server
  is not auto-started.
• tests/e2e/console-errors.spec.ts — opens pages from a PAGES list,
  listens to pageerror / console.error and fails the test on any
  reported error.
• Wiki — wiki/testing/README.md updated with usage instructions and
  caveats.

🔧 Technical Changes

Build & Bundling

• Webpack instead of Turbopack — Turbopack produced invalid client
  chunks (ReferenceError: boolean is not defined) for the
  react-markdown / rehype ecosystem. build:front now runs
  next build --webpack; the Turbopack toggle is gated behind
  NEXT_USE_TURBOPACK=true.
• build:analyze — bundle analyzer script restored on top of the
  Webpack build (@next/bundle-analyzer).

Tool: Shell Execute

• The Build Command node now prepends cd <cwd> && instead of relying on
  child_process cwd option, fixing edge cases where the target shell
  ignored or mis-quoted the working directory.

Docker / Sharp

• Updated the Dockerfile so sharp resolves correctly in production
  images.

📦 Dependencies

• @next/bundle-analyzer ^16.2.6
• @playwright/test ^1.59.1
• es-check (dev)

📚 Documentation

• New wiki/skills/README.md covering layout, naming convention, manifest,
  GraphQL API, executor, and build notes; linked from wiki/README.md.

---

Full Changelog: v1.10.0...v1.11.0

v1.10.0

v1.10.0 — Markdown Editor & Analytics

🚀 New Features

Betterlytics Analytics

Privacy-focused analytics integration using Betterlytics.

• Web Vitals tracking — Core Web Vitals metrics (LCP, FID, CLS)
• Outbound links tracking — Full URL tracking for external links
• Error tracking — JavaScript errors and console errors

Configuration:

NEXT_PUBLIC_BETTERLYTICS_SITE_ID=your-site-id
NEXT_PUBLIC_BETTERLYTICS_SCRIPT_URL=https://your-betterlytics-instance/script.js
NEXT_PUBLIC_BETTERLYTICS_SERVER_URL=https://your-betterlytics-instance

Markdown Editor Improvements

Extended toolbar and editing capabilities:

• Table editing — Insert and edit tables
• Strikethrough, subscript, superscript — Additional text formatting
• Code blocks — Insert code blocks with syntax highlighting
• Source mode — Toggle between rich-text and source view
• Markdown shortcuts — Standard markdown shortcuts support

Markdown Styles

New styles for markdown content:

• Horizontal rules (hr)
• Task lists with checkboxes
• Subscript and superscript
• Footnotes

🔧 Technical Changes

Fetch Request Tool

• Allow internal hosts and IPs (previously blocked)
• Accept headers and body as objects (not just JSON strings)
• Added input parsing node for flexible parameter handling

Agent Improvements

• Pass request params to agent context
• Change default model to google/gemini-3.1-flash-lite-preview

🐛 Bug Fixes

• Fix links color in chat widget (white links on user messages)

📦 Dependencies

• @betterlytics/tracker ^0.2.2

---

Full Changelog: v1.9.0...v1.10.0

v1.9.0 — LLM Client & Knowledge Base

🚀 New Features

LLM Client & GraphQL API

Direct access to LLM capabilities via GraphQL — no n8n workflows required.

Why this matters:

1. Experimentation playground — test message chains, tool calling patterns, and multi-turn conversations directly through GraphQL Playground
2. Server-side integration — use llmClient in resolvers, middleware, and background jobs without workflow overhead
3. Full OpenAI-compatible API — supports vision, tool calls, and conversation history

Mutations:

• llmCompletion — raw text completion
• llmChatCompletion — multi-turn conversations with support for:
  • System/user/assistant roles
  • Vision (image_url content parts)
  • Tool calling history (tool_calls, tool_call_id)

---

Example 1: Tool Calling History

Simulate assistant tool calls and tool responses in conversation:

mutation llmChatCompletionWithToolHistory {
  llmChatCompletion(
    input: {
      messages: [
        { role: system, content: "You are a calculator. Use the calculate function." }
        { role: user, content: "What is 25 * 4?" }
        {
          role: assistant
          content: null
          toolCalls: [
            {
              id: "call_123"
              type: "function"
              function: {
                name: "calculate"
                arguments: "{\"expression\": \"25 * 4\"}"
              }
            }
          ]
        }
        {
          role: tool
          toolCallId: "call_123"
          content: "100"
        }
      ]
    }
  ) {
    text
    finishReason
    usage { promptTokens completionTokens totalTokens }
  }
}

---

Example 2: Multi-User Chat Context

Test how the model tracks multiple participants:

mutation llmChatCompletionMultiUserContext {
  llmChatCompletion(
    input: {
      messages: [
        { role: system, content: "You observe a group chat. Track the conversation." }
        { role: user, content: "[Alex]: I finished the ML project today." }
        { role: assistant, content: "Noted: Alex finished ML project." }
        { role: user, content: "[Maria]: Nice! I'm still debugging the database issue." }
        { role: assistant, content: "Noted: Maria has a DB bug." }
        { role: user, content: "[Alex]: Maria, what's the error?" }
        { role: user, content: "[Maria]: Index corruption after migration." }
        { role: user, content: "Questions:\n1. Who finished a project?\n2. What is Maria's problem?\n3. How many people are in the chat?" }
      ]
    }
  ) {
    text
    finishReason
  }
}

---

Example 3: Chain-of-Thought Reasoning

Guide the model through step-by-step analysis:

mutation llmChainOfThought {
  llmChatCompletion(
    input: {
      messages: [
        { 
          role: system
          content: "You are a code reviewer. Analyze code step by step: 1) Identify the purpose, 2) Find potential bugs, 3) Suggest improvements."
        }
        { 
          role: user
          content: "Review this function:\n\nfunction divide(a, b) {\n  return a / b;\n}"
        }
      ]
      temperature: 0.3
    }
  ) {
    text
    finishReason
  }
}

---

Example 4: Summarization Pipeline

Build document processing chains:

mutation llmSummarize {
  llmChatCompletion(
    input: {
      messages: [
        { role: system, content: "Summarize text in 2-3 sentences. Be concise." }
        { role: user, content: "The quick brown fox jumps over the lazy dog. This sentence contains every letter of the English alphabet and is commonly used for font testing and typing practice. It was first used by Western Union in the late 1800s to test telegraph equipment." }
        { role: assistant, content: "A pangram sentence containing all alphabet letters, used since the 1800s for testing telegraph equipment and later for font/typing tests." }
        { role: user, content: "Now summarize: GraphQL is a query language for APIs that gives clients the power to ask for exactly what they need. It was developed by Facebook in 2012 and open-sourced in 2015. Unlike REST, GraphQL uses a single endpoint and allows clients to specify the structure of the response." }
      ]
      temperature: 0.2
    }
  ) {
    text
    finishReason
  }
}

---

Server-Side Usage

Use llmClient directly in resolvers for automated processing:

// Auto-generate file metadata from images
async function processFile({ file, ctx }: ProcessFileProps) {
  const { prisma, llmClient } = ctx

  const imageBuffer = await readFile(join('uploads', file.path))
  const base64 = imageBuffer.toString('base64')
  const dataUrl = `data:${file.mimetype};base64,${base64}`

  const response = await llmClient.chatCompletion({
    messages: [
      {
        role: 'user',
        content: [
          { type: 'image_url', image_url: { url: dataUrl } },
          { type: 'text', text: 'Return JSON: { "name": "...", "description": "..." }' },
        ],
      },
    ],
    max_tokens: 500,
    temperature: 0.3,
  })

  const result = JSON.parse(response.choices[0].message.content)
  
  await prisma.file.update({
    where: { id: file.id },
    data: { name: result.name, description: result.description },
  })
}

// Auto-tag content based on text analysis
async function autoTagConcept({ concept, ctx }: AutoTagProps) {
  const { prisma, llmClient } = ctx

  const response = await llmClient.chatCompletion({
    messages: [
      { role: 'system', content: 'Extract 3-5 tags from the text. Return JSON array: ["tag1", "tag2"]' },
      { role: 'user', content: concept.content },
    ],
    temperature: 0.2,
  })

  const tags = JSON.parse(response.choices[0].message.content)
  
  await prisma.kBConcept.update({
    where: { id: concept.id },
    data: { data: { tags } },
  })
}

---

Configuration:

# Add to docker/.env
LLM_API_URL=http://llama:8080/v1

Permissions: Both mutations require isSudo.

File Management Module

Full CRUD operations for files with new frontend pages:

• Files list page with pagination and filtering
• File detail page with view and edit modes
• FileItem component for list and full view variants

New File model fields:

• description — file description
• content — file content
• Unique constraint on path

GraphQL additions:

• FileWhereInput — flexible filtering with where/orderBy params
• updateFile mutation with permissions

Concepts Management

Complete concepts management system:

• Concepts list page with pagination
• Concept detail page with view and edit form
• ConceptItem component for list and full view variants

GraphQL additions:

• concept query for fetching single concept
• conceptsCount and conceptsConnection queries
• KBConceptWhereUniqueInput for single concept queries
• Updated updateConcept mutation with where input and sudo check

KBConceptFile Relation

New many-to-many relation between concepts and files:

• Link files to knowledge base concepts
• Prisma migration for KBConceptFile join table

Concept Hierarchy

Hierarchical structure for knowledge base concepts:

New KBConcept fields:

• parentId — parent concept reference
• rootId — root concept reference
• code — unique concept code
• data — JSON data field

Relations:

• ConceptTree — self-referential parent-child
• ConceptRoot — self-referential root reference

🔧 Technical Changes

GraphQL Filters Refactoring

• Add StringNullableFilter and QueryModeEnum for flexible filtering
• Add buildStringNullableFilterWhere helper for reusable filter logic
• Consolidate enums from separate files into shared types/enums.ts
• Remove redundant enums.ts files from KB entities

Image Resizer Refactoring

Extract helpers into separate files for better maintainability:

server/middleware/imageResizer/
├── helpers/
│   ├── flattenAlphaToBackground.ts
│   ├── parseBackgroundColor.ts
│   └── resizeImg.ts
├── index.ts
└── interfaces.ts

n8n Workflow Improvements

• Add hasKBTools, hasUpdateOwnProfileTool config options
• Inject user profile info into agent context
• Always create graphql request tool workflow

Task Improvements

• Add assigneeId field to Task inputs
• Fix create/update task resolvers for assignee handling

🐛 Bug Fixes

• Fix KB queries with improved buildWhere helper
• Fix getConceptsConnectionQueryVariables in concepts page
• Fix conceptsResolver with proper filtering
• Fix imageResizerMiddleware handling
• Fix prism-js in Markdown editor

📁 New Files

server/
├── llm/
│   └── client/
│       ├── index.ts                    # LLM client implementation
│       └── interfaces.ts               # LLM types
├── schema/types/
│   ├── LLM/
│   │   ├── index.ts                    # LLM GraphQL type
│   │   ├── interfaces.ts               # LLM interfaces
│   │   ├── types.ts                    # LlamaUsage type
│   │   └── resolvers/
│   │       ├── chatCompletion.ts       # Chat completion mutation
│   │       └── completion.ts           # Text completion mutation
│   └── KBConceptFile/
│       └── index.ts                    # Concept-File relation
└── middleware/imageResizer/helpers/
    ├── flattenAlphaToBackground.ts
    ├── parseBackgroundColor.ts
    └── resizeImg.ts

src/components/pages/
├── Files/
│   ├── index.tsx                       # Files list page
│   ├── interfaces.ts
│   ├── helpers/index.ts
│   ├── View/
│   │   ├── index.tsx
│   │   ├── styles.ts
│   │   └── FileItem/
│   │       ├── index.tsx
│   │       ├── interfaces.ts
│   │       ├── st...

v1.8.0 — Computer Vision

🚀 New Features

Local Image Recognition

The agent can now analyze images locally using Qwen3.5 vision model via llama.cpp. No cloud APIs required — all processing happens on your GPU.

Capabilities:

• Describe images, charts, diagrams
• Analyze mathematical graphs and functions
• Read text from images (OCR)
• Answer questions about visual content

Configuration:

# Main model (4B recommended for vision tasks)
LLAMA_MODEL=unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf

# Vision encoder (required for image input)
LLAMA_MMPROJ=unsloth/Qwen3.5-4B-GGUF/mmproj-F16.gguf

Both files are auto-downloaded from HuggingFace on first run (~5 GB total).

OpenAI-Compatible Vision API

Send images via the standard /v1/chat/completions endpoint:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,YOUR_BASE64"}},
        {"type": "text", "text": "What is in this image?"}
      ]
    }],
    "max_tokens": 1000
  }'

Note: Images must be sent as base64 — llama.cpp doesn't fetch URLs directly.

📚 Documentation

Computer Vision Guide

New documentation for vision capabilities:

• Configuration and setup
• API usage examples
• Troubleshooting common issues
• Performance benchmarks

→ See wiki/computer-vision/README.md

Updated llama-server Documentation

• Added LLAMA_MMPROJ environment variable
• Added LLAMA_CTX_SIZE and LLAMA_VERBOSE options
• Updated model recommendations for vision tasks

🔧 Technical Changes

entrypoint.sh Enhancements

The llama.cpp entrypoint script now handles multimodal projection files:

• Parses LLAMA_MMPROJ environment variable (format: owner/repo/filename)
• Auto-downloads mmproj from HuggingFace if not present
• Passes --mmproj flag to llama-server when configured

docker-compose.yml Updates

• Added LLAMA_MMPROJ environment variable to llama service
• Added LLAMA_CTX_SIZE for context window configuration
• Added LLAMA_VERBOSE for debug logging

💡 Tips

Transparent PNG Handling

Vision models expect RGB images. PNG files with transparency may render incorrectly (transparent areas become black).

Solution: Add a white background before sending:

# Using ImageMagick
convert input.png -background white -flatten output.png

Recommended Settings

Use Case	Model	VRAM
Testing	Qwen3.5-0.8B	~2 GB
Production	Qwen3.5-4B + mmproj	~6 GB
High quality	Qwen3.5-4B-BF16 + mmproj	~10 GB

Performance

On NVIDIA RTX 3060/4060:

• Image encoding: ~100-300ms
• Text generation: ~40 tokens/sec

📁 New Files

wiki/
└── computer-vision/
    └── README.md                    # Vision setup and usage guide

docker/
└── llama/
    └── entrypoint.sh                # Updated with mmproj support

⚙️ Configuration

Add to docker/.env:

# Vision model (recommended)
LLAMA_MODEL=unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-Q8_0.gguf
LLAMA_MMPROJ=unsloth/Qwen3.5-4B-GGUF/mmproj-F16.gguf

# Optional: increase context for large images
LLAMA_CTX_SIZE=8192

# Optional: enable verbose logging
LLAMA_VERBOSE=1

Restart llama service after configuration:

cd docker
docker compose -f docker-compose.yml -f docker-compose.dev.yml up -d llama

---

Full Changelog: v1.7.0...v1.8.0

v1.7.0 — Agent Identity & Security

🚀 New Features

Agent Self-Profile Management

The agent now has full awareness of its own identity and can update its profile autonomously.

What's new:

• Agent receives its profile data (fullname, intro, content) in context
• New update_profile tool allows the agent to modify its own profile
• Built-in safety instructions protect against identity manipulation (prompt injection)

Tool capabilities:

• Update fullname, intro, content fields
• Reasoning required before any profile change
• Agent maintains identity integrity across sessions

Custom System Message via Credentials

System message can now be configured per agent instance through credentials file (gitignored):

{
  "agentName": "Chat Agent",
  "systemMessage": "Custom system message for this agent instance"
}

Benefits:

• Individual agent tuning without repository changes
• Credentials are gitignored — safe for sensitive instructions
• Override default system message from code

Password Management UI

Administrators can now change user passwords directly from the admin interface.

• New PasswordField component with visibility toggle
• updateUser mutation accepts password field
• Password properly hashed before storage

🔐 Security — Breaking Change

Unified JWT Token System

⚠️ BREAKING: All existing auth tokens are invalidated.

Complete overhaul of the authentication system:

Before	After
JWT-only verification	Database-backed verification
No token revocation	Revocation support via DB
Implicit token types	Explicit TokenType enum (Auth, Referrer)
Sync verification	Async with Prisma dependency

What changed:

• TokenType enum for explicit token classification
• Tokens stored in Token table with expiredAt field
• verifyToken is now async with database lookup
• All auth resolvers updated (signin, signup, authEthAccount, authViaTelegram)

Benefits:

• Token revocation (logout from all devices)
• Audit trail of all issued tokens
• Type-safe token handling
• Foundation for multi-device session management

📚 Documentation

Agent Philosophy

New documentation explaining the core ideology:

• What You Get — autonomous entity, not a chatbot
• Agent as First-Class Participant — comparison with traditional assistants
• Core Principles — Transparency, Knowledge, Identity, Autonomy, Local-first
• Experience System — reflexes and behavioral learning
• Agent World — knowledge graph with attention mechanics

→ See wiki/agent-philosophy.md

Updated README

Rewritten introduction emphasizing the unique value proposition:

• Autonomous AI entity with own identity and credentials
• Epistemic knowledge with confidence tracking
• Mandatory reasoning for all actions

🔧 Technical Changes

System Message Refactoring

• systemMessagePath replaced with systemMessage directly
• System message reading logic moved to individual workflows
• Removed base-system-message.md from agent-factory
• More flexible — system message can be built programmatically

📁 New Files

wiki/
└── agent-philosophy.md           # Core ideology documentation

server/n8n/workflows/
├── agent-chat/nodes/
│   ├── KB/updateProfile/         # Profile update workflow
│   └── execTool/tools/
│       └── updateProfile.ts      # Profile update tool
└── helpers/
    └── objectToMarkdown.ts       # Helper for formatting

src/ui-kit/controls/
└── PasswordField/                # Password input component

⚙️ Migration

1. All users must re-authenticate — existing tokens are invalid
2. No database migration required — Token table already exists
3. Update agent credentials if you want custom system messages

---

Full Changelog: v1.6.0...v1.7.0

v1.6.0 — Shell Execute Tool

🚀 New Features

Shell Command Execution

Agent can now execute shell commands directly on the host system. This enables autonomous system administration, file operations, and integration with external CLI tools.

Tool capabilities:

• Execute any shell command with arguments
• Specify working directory (cwd)
• Returns stdout, stderr, and exit code
• Full reasoning logging before execution

Usage example (agent perspective):

Tool: shell_execute
Command: git status
Cwd: /app

Security

Shell execution is controlled by the canExecuteShell flag in AgentFactoryConfig. Disabled by default — must be explicitly enabled per agent.

const config: AgentFactoryConfig = {
  // ...
  canExecuteShell: true,
}

🐳 Docker Improvements

Additional Utilities

The app container now includes essential CLI tools for agent operations:

• git — version control
• jq — JSON processing
• mc — Midnight Commander file manager
• nmap — network exploration
• sudo — privileged operations

Traefik Routing

Added dedicated route for n8n UI at n8n.localhost for easier workflow debugging during development.

🔧 Build Configuration

Storage Directory Exclusion

The storage/ directory is now properly excluded from:

• TypeScript compilation (tsconfig.json)
• ESLint checking (eslint.config.cjs)
• Prettier formatting (.prettierignore)

This prevents build errors when storage contains non-TypeScript files or generated content.

📁 New Files

server/n8n/workflows/
├── tool-shell-execute/
│   └── index.ts              # Shell Execute workflow
└── agent-chat/nodes/execTool/tools/
    └── shellExecute.ts       # Tool node definition

⚙️ Configuration

To enable shell execution for an agent, add to agent config:

canExecuteShell: true

The tool will appear in the agent's available tools and can be called with:

• command — shell command to execute (required)
• cwd — working directory (optional, defaults to /)
• reasoning — explanation of why this command is needed (required by ExecTool pattern)

---

Full Changelog: v1.5.0...v1.6.0

Added local llama.cpp

v1.5.0 — Local LLM Server with CUDA Support

🚀 New Features

Local llama.cpp Server

Run AI models locally without external API dependencies. The new llama service is included in docker-compose with full NVIDIA CUDA support.

Key capabilities:

• Auto-download models from HuggingFace on first run
• OpenAI-compatible API (/v1/chat/completions, /v1/completions)
• Configurable via environment variables
• Support for HuggingFace token for gated models

Default model: unsloth/Qwen3.5-0.8B-GGUF/Qwen3.5-0.8B-Q8_0.gguf

⚙️ Configuration

Add to docker/.env:

LLAMA_MODEL=unsloth/Qwen3.5-0.8B-GGUF/Qwen3.5-0.8B-Q8_0.gguf
LLAMA_GPU_LAYERS=99
LLAMA_CTX_SIZE=4096
HUGGINGFACE_TOKEN=hf_xxx...  # optional, for gated models

Configure credentials/system/openrouter.json:

[
  {
    "id": "openrouter-cred",
    "name": "OpenRouter",
    "type": "openRouterApi",
    "data": { "apiKey": "llama", "url": "http://llama:8080/v1" }
  }
]

⚠️ Requirements

• NVIDIA GPU with CUDA support
• CUDA drivers installed on host (nvidia-smi should work)
• Docker with NVIDIA Container Toolkit

📚 Documentation

• Local LLM Server Guide
• Credentials Setup

🏃 Quick Start

cd docker
docker compose -f docker-compose.yml -f docker-compose.dev.yml up llama

---

Full Changelog: v1.4.0...v1.5.0

1.3.0: feat: Posts module, referral registration, sitemap

🗒️ Posts Module
Full CRUD for posts with revision history. Every update to a post creates a PostRevision snapshot preserving previous content, title, description, intro, status and signature.

Status lifecycle: draft → published → unpublished
Comments are implemented as child posts (parentId / rootId).

Validation: a validatePost query provides pre-flight checks with hard limits (title ≤ 512 chars, description ≤ 3072 chars) and SEO warnings.

AI guideline: a postGuidelines query returns a Markdown guide describing post structure and signing flow — useful for AI agents composing posts.

✍️ Post Signing (MetaMask)
Posts can be cryptographically signed to verify authorship:

Call getPostSignData(postId) — returns a canonical JSON message and a 5-minute server token.
Sign the message with your Ethereum private key (MetaMask).
Call signPost(postId, serverToken, userSignature) — server verifies both tokens and stores the signature.
Signature is invalidated on any subsequent post update.

🔐 Referral Registration
Registration now requires a referral token by default.

Existing users generate tokens via createReferrerToken mutation (JWT, default TTL: 1 hour).
Token can be delivered via URL: ?referrerToken=.
New users are linked to their referrer (referrerId field on User).
Bypass: MetaMask and Telegram auth skip the token requirement.
Set NEXT_PUBLIC_SITE_SIGNUP_STRATEGY=ANY to disable the requirement entirely.

👤 User Statuses
New newbie status added to UserStatus enum.

Status Create posts Comment Notes
active ✅ ✅ Full access
newbie ❌ ✅ Default for new users (configurable)
blocked ❌ ❌ Suspended
Set default status via USER_DEFAULT_STATUS env var.
Admin can toggle user status via new updateUser mutation (sudo only).

🗺️ Sitemap Generation
Automatic XML sitemaps:

/sitemap.xml — sitemap index
/sitemap/main.xml — main pages
/sitemap/posts.xml — published posts only
/sitemap/users.xml — active users only

🌱 Prisma Seed
A seed script now runs automatically on deploy (npm run prisma:seed).
Set SUDO_PASSWORD to create an admin user with full sudo access.

⚙️ Feature Flags
New environment variables to control which sections are enabled:

Variable Effect
NEXT_PUBLIC_POSTS_ENABLED Show/hide Posts in sidebar
NEXT_PUBLIC_WORLD3D_ENABLED Show/hide Metaverse in sidebar
NEXT_PUBLIC_CRYPTO_ENABLED Show/hide Balance, Transactions, Send Transfer
NEXT_PUBLIC_SITE_TITLE Global site title
NEXT_PUBLIC_MAIN_PAGE_TITLE Main page title

🛠️ Other Changes
JWT helper extracted to server/helpers/jwt.ts (removed code duplication across 3 files)
useBoolean hook — ergonomic boolean state with setTrue, setFalse, toggle
useCopy hook — clipboard copy with snackbar feedback
SeparatorStyled — flex spacer component
buildPostWhere / buildUserWhere — centralized query filtering with per-role visibility rules
Users page — added registration date, status badge, pagination
User profile — SEO noindex/nofollow for non-active users
SeoHeaders — always renders robots meta tag; description now accepts null
Telegram widget — conditionally rendered only when NEXT_PUBLIC_TELEGRAM_BOT_NAME is set