Roadmap¶
AEGIS is built in strictly sequential phases — each one must meet defined exit criteria before the next begins, and each ends with a git tag. Phases 0–10 lead to v1.0.0; Phases 11–18 take it from native apps through a full knowledge OS, live integrations, background intelligence, and a complete workspace. Nothing is skipped; nothing ships until it's complete.
Phase status¶
| Phase | Name | Status | Tag |
|---|---|---|---|
| 0 | Governance & Repo Foundation | ✅ Complete | v0.1.0-governance |
| 1 | Runtime Baseline (Odysseus Adoption) | ✅ Complete | v0.0.1-base |
| 2 | Control Plane Core | ✅ Complete | v0.2.0-control-plane |
| 3 | Cognitive Layer (Engines + Memory) | ✅ Complete | v0.3.0-cognitive |
| 4 | Execution Graph (DAG Engine) | ✅ Complete | v0.4.0-dag |
| 5 | Agent System | ✅ Complete | v0.5.0-agent-runtime |
| 6 | Shared Services + API | ✅ Complete | v0.6.0-services |
| 7 | Module System + AI Skills | ✅ Complete | v0.7.0-modules |
| 8 | UI (Web App Foundation) | ✅ Complete | v0.8.0-ui |
| 9 | Ecosystem / Marketplace | ✅ Complete | v0.9.0-marketplace |
| 10 | Production Hardening | ✅ Complete | v1.0.0 |
| 11 | Desktop + Mobile Apps | ✅ Complete | v1.1.0-apps |
| 12 | Public Website + Distribution | 🔄 Current | v1.2.0-public |
| 13 | Knowledge Foundation | ⏳ Planned | v1.3.0-knowledge |
| 14 | Obsidian Bridge | ⏳ Planned | v1.4.0-obsidian |
| 15 | Platform Polish | ⏳ Planned | v1.5.0-polish |
| 16 | Connected Services | ⏳ Planned | v1.6.0-connected |
| 17 | Background Intelligence | ⏳ Planned | v1.7.0-agentic |
| 18 | Workspace & Projects | ⏳ Planned | v1.8.0-workspace |
What's been built¶
Phase 0 — Governance & Repo Foundation The constitutional layer: a mission, vision, architecture constitution, guiding principles, and the ADR system. Structure before code.
Phase 1 — Runtime Baseline AEGIS adopted Odysseus as its initial runtime — wrapping it externally over a network boundary, never vendoring its code. Clean separation from day one.
Phase 2 — Control Plane Core The central brain is live: Orchestrator, Service Registry, Permission System, Module Loader, Tool Manager, Event Bus, Config, and Logging. Every subsequent layer builds on this foundation.
Phase 3 — Cognitive Layer The Cognitive Engine Interface now speaks to Anthropic, OpenAI, Ollama, and Odysseus — each engine is one adapter among several. Hardware-aware capability scoring, multi-layer memory (identity, goals, projects, relationships), knowledge graph, and safety layer are all implemented and tested.
Phase 3 was the architectural pivot. After it, no single AI engine is required. AEGIS routes work to whatever is best available for the task.
Phase 4 — Execution Graph (DAG Engine) The deterministic workflow engine: DAG compiler, executor, persistent SQLite state, and task resumption. All non-trivial execution runs as a compiled graph that can be paused and resumed.
Phase 5 — Agent System Autonomous agents operating on the DAG engine: base agent framework, eight role agents (Memory, Planner, Researcher, Teacher, Reviewer, Coach, Project Manager, Creativity), agent-to-agent messaging, and agents as DAG nodes.
Phase 6 — Shared Services + API Eight shared services — Document Vault, OCR, Knowledge Base (full-text search), Semantic Search, Notifications, Calendar, API Connectors, and universal Token Usage Tracking — plus the authenticated HTTP API gateway with telemetry. The reference flow works end-to-end: a file dropped in the vault is text-extracted, indexed, and searchable by any module.
Phase 7 — Module System + AI Skills Installable, permission-gated modules with validated manifests and dynamic loading; module and add-on templates; module-creation skills; and HEPHAESTUS — the guided module builder. Answer a few questions (or point it at a description file) and it generates, validates, and installs a standard module with no code written. Verified by a non-coder building a working module end-to-end.
Phase 8 — UI (Web App Foundation) A React web app served by the API gateway: streaming agentic chat (the model can use module tools), effort levels (reasoning budget), a model picker across cloud/Ollama/local GGUF engines, persistent conversations, memory management, live logs and task views, and cost/context controls — a daily cloud-token budget with a UI slider, plus context and model-window gauges. Odysseus became fully optional here, as scheduled.
Phase 9 — Ecosystem / Marketplace The module ecosystem: a registry (its own repo — private for now, public when the ecosystem opens) that AEGIS browses from a Marketplace panel; consent-gated installs — every module shows a plain-language review of what it wants, downloads are fingerprint-verified against the registry, and code is scanned before you approve; and sandboxed execution — marketplace modules run in isolated worker processes with no access to secrets, the network, or files outside their own folders.
Phase 10 — Production Hardening
The path to v1.0.0, delivered: CI on every change (tests, compile gate, web-app build, and a Docker image smoke test), request-ID tracing with a live metrics panel, containerized deployment with hardware capability profiles — the same AEGIS kernel runs on a Raspberry Pi, a desktop, or a GPU server, with services enabled to match the detected hardware — the 9-session Interview System that fills AEGIS's long-term memory, deterministic slash commands, and a final security and reliability review.
Phase 11 — Desktop + Mobile Apps
v1.1.0-apps. The web app wrapped as a native Windows/Linux desktop app (Tauri) and an Android app (Capacitor). A unified connect screen — Server + API Key fields, auto-detection via mDNS (aegis.local) and DNS, accepts bare hostnames and Tailscale names. Whisper STT + Piper TTS voice: push-to-talk dictation, hands-free wake word (Vosk by default; Porcupine optional), follow-up conversation window, spoken stop commands. Memory Fabric: entities auto-extracted, memory-to-memory links with lineage, about-time, recall strength, pin, and confidence. Animated presence avatar on the Dashboard; minimap scrollbar; vault folders with source links; customizable dashboard; CI desktop-apps workflow building signed artifacts.
What's next¶
Phase 12 — Public Website + Distribution A public home: landing page, full user manual, downloads, and a setup guide a non-developer can follow from zero to first prompt.
Phase 13 — Knowledge Foundation
Replace the Phase 3 data stack with a production-grade knowledge engine. PostgreSQL + pgvector replaces SQLite + Chroma (profile-gated — Edge keeps SQLite). A knowledge/ vault of Markdown files becomes canonical; an indexer service watches for changes and syncs them into the database within 5 seconds. Hybrid retrieval combines semantic, full-text, tag, relationship, and temporal signals into a single re-ranked result — every feature that reads memory gets better automatically.
Phase 14 — Obsidian Bridge Connect AEGIS to Obsidian — the user's knowledge editor — so notes written there become AEGIS memory automatically. A community Obsidian plugin (separate TypeScript repo) adds real-time sync, an "Ask AEGIS" panel inside the editor, and a semantic relationship sidebar. Two optional modules ship alongside: Memory Canvas (interactive knowledge graph in the AEGIS web UI) and Knowledge Analytics (orphan rate, staleness, coverage gaps) — both installable via the Marketplace.
Phase 15 — Platform Polish
Clear the approved backlog. Module orchestrators: one natural-language command per module (/assistant book dentist Tuesday). Everyday UX: Android back-button navigation, drag-reorder for all sidebar items, Windows tray server (shield icon, no terminal window). AEGIS Mini: a floating avatar overlay on Android — draggable, always available, tap to open a compact chat over any app. Multi-calendar engine: the calendar service expands to support multiple named calendars (offline-first, always available without internet) with 15 selectable calendar types. Five alternative calendars: Enochic (exactly as described in 1 Enoch, 12 months × 30/30/31 per quarter, 4 intercalary days, 364-day perpetual year), IFC/Cotsworth (13 × 28 days + Year Day), Ethiopian Ge'ez (12 × 30 + Pagumē), Hebrew lunisolar, and Islamic Hijri. Ten ancient calendars: Ancient Egyptian (12 × 30 + 5 epagomenal days, direct ancestor of the Gregorian system), Julian (still used by Orthodox churches, currently 13 days behind Gregorian), Mayan (Haab civil + Tzolkin sacred, displayed simultaneously — the 52-year Calendar Round), Babylonian/Mesopotamian (lunisolar ancestor of the Hebrew calendar, the calendar of the Babylonian Exile), Roman pre-Julian (the original 10-month calendar that explains why September is the 9th month), Chinese Traditional (lunisolar with 24 solar terms and the 60-year stem-branch cycle), Hindu/Vedic (Panchanga — five-element date system still in daily use), Greek/Attic (Athenian lunisolar, year starts after the summer solstice, essential for classical history), Celtic/Coligny (recovered from the 2nd-century CE Coligny bronze tablet, most complete ancient Celtic document), and Aztec (Xiuhpohualli civil + Tonalpohualli ritual, parallel to the Mayan system). A Calendar tab in the AEGIS web UI lets you view any calendar type and switch between them.
Phase 16 — Connected Services Give AEGIS eyes on the web and hands in the user's real digital life. A Web Search module registers itself as a service so any other module can look things up — results cited inline. Google Workspace module: Calendar, Gmail, Drive, Contacts. Microsoft 365 module: Outlook Calendar, Outlook Mail, OneDrive. All send/write actions go through an explicit consent gate. All three are optional modules, not hard dependencies.
Phase 17 — Background Intelligence AEGIS works while you're not watching. Scheduled agents run on a cron-style timer (daily brief at 8am, no user action needed). Event-driven triggers fire when something happens (new email → summarise + tag; calendar event approaching → brief). Long-running research tasks execute in the background and push a notification to Android when done. A Proactive Insights module (optional) watches for usage patterns and suggests automations — the scheduler engine works without it.
Phase 18 — Workspace & Projects Materialise the full HX vision. Projects become real workspace containers — conversations, documents, files, knowledge nodes, agents, and tasks all scoped together. Outputs become first-class versioned artifacts (download, version history, export). A collapsible right context panel shows exactly what AEGIS used to produce each reply: which memories, which agents, which tools, which engine. Multi-agent tasks show individual agent contributions in the task view.
The runtime independence thread¶
A thread running through the entire roadmap: AEGIS starts with Odysseus as its initial runtime dependency. Phase by phase, every capability that AEGIS borrows from Odysseus gets replaced by an AEGIS-native implementation. By Phase 8, Odysseus is fully optional. By Phase 10, AEGIS runs without any external runtime dependency at all.
This is documented in ADR-0003: Odysseus is scaffolding, not a foundation.