Roadmap
Phase 1 — Foundation ✅ Live
The current, production-deployed state of Vespera.
| Component | Status |
|---|---|
| Hybrid Model Orchestration (Groq + Gemini) | ✅ Complete |
| Truth Block Protocol (RAG-Lite) | ✅ Complete |
| Cloud Engine — AI Advisor + RAG (186 practices) | ✅ Complete |
| Cloud Engine — Terraform Validation Pipeline | ✅ Complete |
| Cloud Engine — Guardrails (Budget / Security / Compliance) | ✅ Complete |
| D&D — Action Economy Validator | ✅ Complete |
| D&D — Rulebook Ingestor (SRD 2024) | ✅ Complete |
| D&D — Generational Void Cycle (3-Phase Campaign Engine) | ✅ Complete |
| Utility — Translator (4 styles + smart glossaries) | ✅ Complete |
| Utility — TL;DR (chunked + string interned) | ✅ Complete |
| Utility — Moderator (AI hostility scoring + LRU cache) | ✅ Complete |
| Aggressive Memory Hygiene (1GB RAM budget) | ✅ Complete |
| Chain-of-Thought Enforcement | ✅ Complete |
| VesperaPersonality Singleton | ✅ Complete |
Phase 2 — Agentic Dissection 🔨 June 2026
Goal: Break the monolith. Implement self-correction via Sequential Agentic Workflow.
| Feature | Description |
|---|---|
| The Blackboard | Deploy agent_tasks, ai_response_cache, system_corrections SQLite tables. Commands write to DB instead of calling AI directly. |
| Self-Correction | Arbiter (Router) → Weaver-Stylist (Actor) → Deep Critic (Reviewer) pipeline with automatic retries on rule violations. |
| XAI Traceability | logic_trace column on every agent task. /why command lets users see which agent made which decision. |
| Loop Guards | Hard retry_count cap (max 3). Semaphore limiting 3 concurrent pipelines, queue depth 10. |
| Security Mitigations | Cache poisoning prevention, loop injection cap, structured output validation by the Critic. |
See the full technical design: Planned — Lite MAS
Phase 3 — Modular Skill-Set 🔭 July – August 2026
Goal: True scalability and industry-standard tooling integration.
| Feature | Description |
|---|---|
| Skill-Based Architecture | Move logic out of Cogs into standalone skills/ modules. skills/cloud/ holds Terraform templates and CLI handlers. skills/dnd/ holds rulebook fragments and combat logic. |
| Model Context Protocol (MCP) | Integrate an MCP client, allowing Vespera to interface with any MCP-compatible external tool without writing custom wrappers per API. |
| The Proactive Pulse | A background asyncio task that wakes on a configurable schedule (e.g., every 60 minutes) to proactively audit cloud FinOps costs or check narrative progression — making Vespera an agent, not just a responder. |
| Semantic Long-Term Memory (RAG 2.0) | Upgrade from keyword RAG to embedding-based semantic retrieval, allowing contextual memory that persists meaningfully across campaign sessions or cloud projects. |
Future Possibilities
These are exploratory ideas with no committed timeline:
- Auto-Improvement Cron: A nightly background job that reviews Critic rejection logs and proposes prompt refinements
- Persona Guardrails: Formal constitutional constraints on Vespera's personality to prevent tone drift even under adversarial prompting
- Cross-Guild Knowledge Sharing: Opt-in anonymized best-practice sharing between servers for the Cloud Engine RAG base