Quilltap v2.7.0 Release Notes

Highlights

File Storage Abstraction - Pluggable storage backends with mount points. Configure multiple storage locations, migrate files between them, and use S3 or local filesystem interchangeably.

Context Compression - Sliding window compression for long conversations using a cheap LLM. Keeps recent messages intact while summarizing older ones, dramatically reducing token costs.

Project-Based Organization - New Projects feature for organizing chats, files, and characters with optional instructions injected into system prompts.

MCP Integration - Model Context Protocol server connector plugin enables connecting to external MCP servers and exposing their tools to Claude.

Homepage Redesign - Three-column responsive layout with welcome section, quick actions, recent chats, projects, and favorite characters.

Major Features

File Storage Abstraction System

New generic FileStorageBackend interface for pluggable storage providers
Built-in local filesystem backend (configurable via QUILLTAP_FILE_STORAGE_PATH)
S3 backend moved to optional plugin (qtap-plugin-storage-s3)
Mount point system for configuring multiple storage locations in Settings → File Storage
Files can be stored on different backends per mount point
Per-project mount point assignment with batch file migration when changing
Encrypted secrets storage for backend credentials (AES-256-GCM)
Orphan file recovery tool: scan backends for untracked files and adopt them
First-class folder entities with database persistence (empty folders now survive)
Project-based S3 key structure: users/{userId}/{projectId}/{folderPath}/{fileId}_{filename}

Sliding Window Context Compression

Compresses older messages using a cheap LLM while keeping last N messages (default 5) in full context
Also compresses system prompts when compression is active (tool definitions never compressed)
New request_full_context tool allows AI to reload full conversation if needed
Async pre-compression: starts immediately after receiving response, cached for next message
Configurable window size (3–10 messages) and target token counts in Chat Settings
Keep-alive pings during compression prevent proxy/ALB timeouts

Projects Feature

Organize chats, files, and characters under Projects with optional instructions
Project instructions injected into system prompts for all associated chats
Character roster management with "allow any character" option
project_info LLM tool for accessing project context (get_info, get_instructions, list_files, read_file, search_files)
Project detail page with expandable cards: Files, Characters, Settings
Sidebar integration with collapsible projects showing nested chats
Full backup/restore and sync support for projects

Homepage Redesign

Three-column responsive layout (3 columns desktop, 2 tablet, 1 mobile)
WelcomeSection with personalized greeting
QuickActionsRow: Start Chat, Continue Last, New Project, Generate Image
RecentChatsSection with avatars, message counts, and quick-hide filtering
ProjectsSection showing 4 most recent active projects
CharactersSection with 2×2 grid of favorites and quick chat buttons
Standalone /generate-image page for image generation outside chats

MCP Server Connector Plugin

New built-in qtap-plugin-mcp for connecting to Model Context Protocol servers
Uses official @modelcontextprotocol/sdk with Streamable HTTP and SSE transports
Dynamically discovers and exposes tools from connected MCP servers
Collision-aware naming: tools use original names, only prefixed on collision
Supports multiple simultaneous server connections
Authentication: Bearer tokens, API keys, and custom headers
Auto-reconnection with configurable retry attempts
Configuration via Settings → Tools with JSON array of server configs

Multi-Tool Plugin Support

New getToolDefinitions(config) and executeByName(toolName, input, context) pattern
All tool plugins standardized on multi-tool pattern for consistency
Tool registry handles both static and dynamic (MCP) tools
Dynamic tool discovery happens at request time, not startup
Updated @quilltap/plugin-types to v1.9.0

API & Backend

V1 API Migration (Complete)

All legacy /api/* routes now return 410 Gone with redirect instructions
New /api/v1/ namespace with action dispatch pattern (?action=)
Consolidated endpoints for all entity types with consistent response formats
Authentication routes: /api/v1/auth/signup, /api/v1/auth/change-password, /api/v1/auth/delete-account, /api/v1/auth/2fa/*, /api/v1/auth/oauth/[provider]/*
System routes: /api/v1/system/deployment, /api/v1/system/plugins/initialize, /api/v1/system/backup, /api/v1/system/restore, /api/v1/system/mount-points, /api/v1/system/tools
All frontend components updated to use v1 endpoints

Migration System Overhaul

Migrations now run in instrumentation.ts BEFORE server accepts requests
If migrations fail, process exits with code 1 (container won't start with incompatible data)
Removed qtap-plugin-upgrade plugin; migrations now in migrations/ directory
MigrationRunner class handles dependency sorting, state tracking, and execution
Migration state stored in MongoDB migrations_state collection

Token Usage Tracking

Added promptTokens, completionTokens fields to messages
Added aggregate fields to chats: totalPromptTokens, totalCompletionTokens, estimatedCostUSD
Added usage tracking to connection profiles: totalTokens, messageCount
New SystemEvent type for tracking cheap LLM operations
Cost estimation using OpenRouter pricing when available, with fallback
Token display settings: per-message tokens/cost, chat totals, system events
New UI components: TokenBadge, SystemEventMessage, ChatCostSummary

Chat Improvements

File Management LLM Tool

New file_management tool for LLM access to project and general files
Operations: list, read, write, and organize files with folder support
Write operations require user permission (per-project or general scope)
Inline file write permission prompt in chat with quick approve/deny
Deferred execution: tool pauses for approval, LLM sees "Waiting for user approval"
Permission schema with SINGLE_FILE, PROJECT, GENERAL scopes

Enhanced First Message Context

Characters receive relevant context when speaking first in new chats
Includes recent and semantically-relevant memories about other participants (3–5 per participant)
Project name, description, and instructions included if chat is in a project
Generic greeting fallback when LLM content filter blocks auto-generation

Graceful Error Recovery

When requests exceed LLM limits, system attempts to recover gracefully
Sends simplified message to LLM explaining what happened with attachment details
LLM provides in-character response suggesting alternatives
Two-tier fallback: LLM-generated recovery, then static fallback
Recovery messages saved with recoveryType metadata field
Supports: token limits, PDF page limits (max 100), image/file size limits

Turn Manager Improvements

Chat title moved to page toolbar header (updates automatically on rename)
Manual chat rename with auto-rename toggle in tool palette
Dynamic browser tab title shows chat name
Improved continue mode error handling with better error extraction

File System Improvements

Redesigned File Browser

Grid view with image thumbnails (on-demand generation with S3 caching)
List view with sortable columns (Name, Associations, Type, Date)
Preview modal for images, PDFs (via PDF.js), text/code files
Syntax highlighting for code files using react-syntax-highlighter (30+ languages)
Markdown files render with formatting, YAML frontmatter in "Document Info" section
Wikilinks supported: [[File]], [[File#Header]], [[File|Text]]
Copy button for text file preview
Folder operations: create, rename, delete
File operations: move to folder, rename, delete with association management

File Upload Improvements

Removed file type restrictions; all types can now be uploaded
Backend automatically detects plain text content by sampling bytes
MIME type inferred from file extension when browser provides generic types
Paste images directly into chat textarea (auto-uploaded and attached)
Unicode filenames work correctly (Base64 encoding for S3 metadata)
Automatic image resizing for provider size limits (Anthropic: 5MB, others: 20MB)

Enhanced File Deletion

Files linked to chats/characters show detailed confirmation dialog before deletion
Dialog lists all associated characters and chats using the file
Deleting dissociates from linked entities; messages get deletion note appended
Direct repository updates ensure memories are NOT regenerated from deletion notes

Provider & Tool Improvements

Tool Plugin Development

New TOOL_PROVIDER plugin capability for custom LLM tools
Tool plugins define schemas, validation, execution handlers, and result formatting
Plugin configuration UI for per-user settings via configSchema
PluginConfigModal dynamically renders form fields
Configuration stored per-user in MongoDB
Documented in TOOL_PLUGIN_DEVELOPMENT.md

Curl Tool Plugin

New qtap-plugin-curl provides curl tool for HTTP requests
Supports GET, POST, PUT, DELETE, PATCH, HEAD, OPTIONS
Options: --url, --request, --header, --data, --user-agent, --max-time, --location, --insecure, --render
--render converts HTML responses to plain text
Configurable URL allowlist for security
Blocks private/local IP addresses for SSRF protection

Provider Improvements

Anthropic plugin fetches models from API instead of hardcoded list (v1.0.10)
OpenRouter streaming token tracking fixed
OpenAI-compatible streaming token tracking fixed
GPT-5 reasoning models: increased max_completion_tokens to 4096
OpenRouter public API used for cost estimation fallback
Web search tool decoupled from native provider web search (useNativeWebSearch field)

Theming & CSS

New qt-* Utility Classes

Alert CSS variables: --qt-alert-success-*, --qt-alert-warning-*, --qt-alert-error-*, --qt-alert-info-*
Entity badge CSS variables: character, persona, chat, tag, memory
Status badge CSS variables: enabled, disabled, related, manual, auto
Plugin source badge CSS variables: included, npm, git, manual
Warning button CSS variables: --qt-button-warning-*
Search highlight class: qt-highlight with customizable styling
Filter chip classes: qt-filter-chip, qt-filter-chip-active
File preview classes: qt-file-preview-*, qt-wikilink, qt-wikilink-broken
Copy button classes: qt-copy-button

Theme Updates

Ocean v1.2.6: Copy button variables, file preview styling
Earl Grey v1.2.2: Wikilink styling, file preview panels
Rains v1.1.5: Wikilink styling, file preview panels
theme-storybook v1.0.13: All new CSS variables and component classes

Testing

Comprehensive Unit Test Expansion

Phase 1 (API Layer): 201 tests for responses, middleware, search service
Phase 2 (Chat Services): 97 tests for streaming, tool execution, pseudo-tool parser, conversation builder
Phase 3 (Data Infrastructure): 143 tests for import/export, theme system
Phase 4 (Hooks & Auth): 242 tests for 9 React hooks, Arctic OAuth modules
Phase 5 (Utilities & Services): 161 tests for utility functions, remaining services
Grand Total: 844+ new tests across 40+ modules with 90%+ coverage on critical paths
All 3438 tests now pass (164 test suites)

Feature-Specific Tests

Context compression: 22 tests
Compression cache service: 17 tests
Text detection utility: 47 tests
LLM error types: 47 tests
Recovery service: 32 tests
Tool registry: 28 tests
Orphan recovery: 13 tests
Folders repository: 28 tests
First message context: 18 tests
Memory processor inter-character tracking: 11 tests

Bug Fixes

Critical Fixes

Sign out now properly clears session and resets UI (wrong cookie name)
Startup state persists across Next.js module reloads (moved to global namespace)
All authenticated API routes wait for server startup to complete
Cloud backups appear in backup list after creation (missing metadata entries)
Plugin toggle works with PUT handler for enable/disable
Quick Chat dialog loads connection profiles correctly
Sync push response format returns arrays instead of numbers
MongoDB migrations work in hosted environments without admin access

Chat & Messaging Fixes

Streaming reliability during long context compression operations
Generic greeting fallback when LLM content filter blocks auto-generation
OpenRouter streaming tool calls detected (camelCase handling)
Memory extraction now properly links memories to user-controlled characters
Template variables in user character descriptions properly substituted
File write permission prompt appears correctly

UI Fixes

Recent chats quick-hide filtering on homepage
Character avatar setting from gallery (HTTP method mismatch)
Image tagging in character photo gallery (correct API endpoints)
Plugin source detection showing npm packages correctly
Cancel button in form modals always enabled except during loading
Settings UI data structure and API endpoint alignment

Refactoring

Route Decomposition

Characters route: 821 → 428 lines (extracted character-wizard.service.ts)
Chats/[id] route: 1,876 → 49 lines (extracted handlers/ and actions/ directories)
Extracted 14 Zod validation schemas, helper functions, and type definitions

Service Extraction

handleInterCharacterMemory: Extracts memories one character learns about another
handleContextSummary: Updates running context summaries using cheap LLM
handleTitleUpdate: Evaluates and updates chat titles based on conversation

CSS Migration

Converted Tailwind color classes to qt-* semantic classes throughout
RenameReplaceTab.tsx: Error alerts, info boxes
plugins-tab.tsx: Plugin source badges

Security

Dependency Updates

esbuild 0.19.0/0.20.0/0.24.0 → 0.27.0 across all 14 plugins (CVE-2024-23334)

Documentation

Major README overhaul: repositioned as general-purpose AI chat platform
Roadmap extracted to features/ROADMAP.md
Provider plugin development guide: docs/PROVIDER_PLUGIN_DEVELOPMENT.md
Testing guide: docs/TESTING_GUIDE.md
File LLM access documentation: docs/FILE_LLM_ACCESS.md
Updated CLAUDE.md documentation references

Uh oh!

2.7.0