Changelog

Quoth v3.3.0: Managed Mode & Socket IPC

SaaS-ready managed pipeline, Unix socket query server for sub-millisecond hook latency, and expanded routing coverage.

Managed Mode Pipeline

• SaaS-ready daemon that runs without user API keys
• Cloud processes JUDGE→DISTILL→CONSOLIDATE server-side via POST /api/v1/pipeline/process

CLI Onboarding Wizard

• Interactive node cli.js init with mode detection, API key setup, auto-daemon start

Unix Socket Query Server

• Daemon serves routing + pattern injection via ~/.quoth/daemon.sock
• Hooks query in ~20ms instead of ~200ms cold start

Venice BGE-M3 Embeddings

• Cloud embeddings migrated from voyage-4-lite to Venice.ai BGE-M3 (1024d, cheaper)

Batch Embeddings

• generateEmbeddingBatch() for cost-optimized MiniLM processing

Expanded Routing

• 20+ categories including Spanish (accent-agnostic)
• Questions route to researcher@0.6

Cross-Org Pattern Sharing

• GET/POST /api/v1/patterns/shared for opt-in anonymous pattern exchange

Doc-Updater Cloud

• POST /api/v1/docs/update endpoint for managed-mode doc auto-updates

Improvements

• Pattern lifecycle tightened: 30-day archive for never-exposed, capacity cap 1000→600
• Exposure-based decay only — inactive patterns no longer penalized
• 66 new tests across pipeline-api, embed-batch, routing-v2, doc-update-api, shared-pull
• Doc manifest hash refresh for accurate staleness detection

---

Quoth v3.3.0: Intelligent Self-Learning Pipeline

Major upgrade to the plugin's learning pipeline, pattern quality, and autonomous documentation maintenance.

Context-Enriched Trajectory Capture

Every tool call now captures full conversation context:

• user_intent — last user prompt (from rolling history of 5)
• conversation_context — last 3 user prompts in chronological order
• llm_reasoning — LLM decision signals extracted from tool inputs (Bash description, Agent prompt, Edit diffs)
• tool_input / tool_output — sanitized full I/O (redacts API keys, JWTs, UUIDs, passwords)

Session-Level Batch Distill

Instead of distilling patterns per tool call, Quoth now batches at session level:

• SessionEnd/PreCompact writes a session_summary to the JSONL
• Daemon processes entire sessions in one LLM call → produces 1-3 higher-quality strategy-level patterns
• Stale session detector (every 10min) generates synthetic summaries for orphaned sessions

Improved Pattern Quality

• DISTILL prompt enforces meaningful names — never raw file paths or tool calls
• Pre-insert dedup: embedding cosine ≥ 0.92 or name prefix match ≥ 80%
• CONSOLIDATE fixed: was always falling back to action: 'new' due to markdown code block parsing bug
• Batch-distilled patterns start at confidence 0.55 and get tagged batch-distilled

3-Tier Temporal Decay

Expanded Task Routing

• ~20 pattern groups (up from 8) with intent-first priority
• Spanish language support (Argentine voseo: arreglá, investigá, desplegá...)
• Word boundaries (\b) on English patterns to prevent false positives
• New categories: fix/debug, refactor, config/setup

Autonomous Doc Maintenance

• Content hash tracking (SHA-256) per source file referenced in each doc
• Nightly pipeline at 3am: Phase A (deep consolidation) → Phase B (doc auto-update)
• LLM incremental update of stale docs with automatic version bump (1.0.0 → 1.0.1)
• Git commit + push after updates
• Session-start reports doc changes once (never repeated)

Unified LLM Provider

All daemon pipeline stages now use Kimi K2.5 via Moonshot API:

• JUDGE, DISTILL, DISTILL-BATCH, CONSOLIDATE, doc-updater — single provider
• Removed dependency on claude CLI subprocess for CONSOLIDATE

Pattern Cleanup

• One-time cleanup archived 699 garbage patterns (raw tool calls) + 109 name duplicates
• Active patterns reduced from 1394 → 586, average confidence 0.50 → 0.67

---

Quoth v3.2.0: Context Injection & Plugin System

• Project-specific context injection at session start (quoth-plugin/context/{project}.md)
• 5 built-in skills: genesis, learn, patterns, quoth-help, quoth-init
• .claude-plugin/ manifest for Claude Code plugin system (hooks, commands, agents, MCP)
• Cleanup legacy directories and obsolete migration scripts

---

Security Audit & Test Repairs

• Fixed all 8 security issues from Ruflo scan + deep agent audit
• Repaired 23 broken tests across plugin and SaaS suites
• Added 6 missing API routes
• Fixed Vercel build issues

---

Quoth v3.0: Multi-Agent, Multi-Organization Platform

Major architectural evolution from single-project to multi-organization platform.

New Tables

• organizations — top-level container for agents and projects
• agents — registered agents with lifecycle management
• agent_projects — optional many-to-many agent-project assignment
• curator_log — curator activity tracking

Extended Tables

• documents — added agent_id, visibility (project/shared), tags
• document_proposals — added agent_id, source_instance
• projects — added organization_id

New MCP Tools

• quoth_agent_register — register agent in organization
• quoth_agent_update — update agent profile/status
• quoth_agent_remove — archive or delete agent
• quoth_agent_list — list agents with filters
• quoth_agent_assign_project — assign agent to project
• quoth_agent_unassign_project — remove assignment

Infrastructure

• RLS policies on all new tables (has_org_access, is_org_admin)
• match_shared_documents RPC for cross-project search
• Document templates (agent-profile, decision-record, how-to, project-brief)
• Backward compatible: existing v2.0 projects auto-migrated

---

Quoth v2.0: AI Memory Architecture - Transform from Agentic RAG to AI Memory

This release fundamentally transforms Quoth from a documentation auditor into a true AI Memory system with bidirectional learning.

Core Architecture Changes

• Local-first storage - .quoth/ folder structure persists knowledge across sessions
• quoth-memory subagent - Sonnet-powered memory interface for context queries (~500 token summaries)
• Session logging - All actions logged to .quoth/sessions/{id}/ for knowledge capture
• Knowledge promotion - User-approved transfer: session → local → remote Quoth

New Skills

• /quoth-init - Initialize AI Memory for a project with configuration wizard
• Genesis v3.0 - Added Phase 0 configuration for strictness and knowledge types

Configurable Strictness Levels

New Hooks

• PreToolUse (Edit/Write) - Gate enforcement with configurable strictness
• PostToolUse (Edit/Write/Bash) - Log actions to session folder
• SubagentStart/Stop - Context injection for other subagents (excludes quoth-memory)
• Updated Stop hook - Knowledge promotion prompt at session end

Local Folder Structure

.quoth/
├── config.json         # Strictness, types, gates
├── decisions.md        # Architecture choices
├── patterns.md         # Code patterns
├── errors.md           # Failures and fixes
├── knowledge.md        # General context
└── sessions/{id}/      # Session-scoped logs

Breaking Changes

• Plugin version bumped to 2.0.0
• New folder structure requires /quoth-init for existing projects
• Hook token overhead increased slightly for enhanced features

---

Quoth Plugin for Claude Code - Lightweight hooks and adaptive guidelines

The new Quoth Plugin replaces heavy persona prompts with an efficient, autonomous integration:

• quoth_guidelines tool - One adaptive tool with three modes (code, review, document) replacing three separate personas. Compact (~150 tokens) or full (~500 tokens) guidelines on demand.
• Lightweight hooks - ~60 tokens per session instead of ~750. Gentle hints that strongly suggest, not force.
• Badge enforcement - 🪶 feather badge displays which patterns were applied when Quoth tools are used.
• 92% token reduction - Hooks dropped from ~750 to ~60 tokens. Full guidelines from 6,000 to 500 tokens.

Two-Command Install:

# Add marketplace (one time)
/plugin marketplace add Montinou/quoth-mcp

# Install plugin (includes MCP + hooks + skills)
/plugin install quoth@quoth-marketplace

The plugin bundles the MCP server, lightweight hooks, and the /quoth-genesis skill. For MCP-only usage without hooks, you can still install via claude mcp add.

Production Hardening - Security and reliability improvements

• Rate limiting on MCP endpoints (60 requests/minute)
• Security headers via middleware (CSP, HSTS, X-Frame-Options)
• Input validation and sanitization for all API endpoints
• Public demo endpoint at /api/mcp/public (read-only, no auth required)
• Improved error handling with structured error responses
• Token validation improvements for API keys

Insights Dashboard - Documentation health monitoring

The Analytics page now tracks documentation health with four new features:

• Drift Detection - Timeline view showing when code diverged from docs, with severity levels (critical, warning, info) and one-click resolution
• Health Dashboard - Visual staleness indicators (fresh <14d, aging 14-30d, stale 30-60d, critical >60d) with health score gauge
• Miss Rate Analytics - Track searches returning no results to identify documentation gaps
• Weekly Health Reports - Automated email reports to project admins every Monday at 9:00 AM UTC via GitHub Actions

Genesis v2.0 - Configurable documentation depth

• Added depth levels: minimal (3 docs), standard (5 docs), comprehensive (11 docs)
• Genesis now asks for confirmation before starting
• Incremental document upload (no batching)
• Better error handling and progress feedback

Dashboard Improvements

• New team management interface
• Role-based access control (admin/editor/viewer)
• Invitation system with email notifications
• API key management with expiration

Search Enhancements

• Upgraded to Jina Embeddings v3 (512 dimensions)
• Added Cohere reranking for improved relevance
• Chunk-level search with context snippets
• Performance improvements for large knowledge bases

MCP Protocol Updates

• OAuth 2.0 authentication support
• Dynamic client registration
• SSE transport option for streaming
• Public demo endpoint (read-only)

---

Initial Release

• Core MCP server with 4 tools: search, read, propose, genesis
• Supabase-powered vector storage
• Multi-tenant project support
• Basic proposal workflow