Roadmap
Sidespace uses its own roadmap system internally. What follows is an honest snapshot of where things stand, what is being worked on, and what is still uncertain.
Current State
Sidespace is at v0.1.0, used daily as the primary development environment for its own codebase (and several other projects in the mcfw workspace). The core loop -- project management, Hoshi chat, Squad View terminals, memory persistence, research pipeline -- is shipped and functional.
The Five Agents
The agent roster is the backbone of the product. Each agent has a specific role and a specific model, chosen to balance capability against cost:
| Agent | Role | Model | Status |
|---|---|---|---|
| Hoshi | Primary assistant. Manages projects, answers questions, executes tasks through 56 MCP tools. | Gemini 3.1 Pro | Shipped. Gemini engine is active, legacy Rust+Sonnet preserved as fallback. |
| Claude Code | Hands-on coding. Writes, edits, and debugs code in live terminal sessions. | Claude (via CLI) | Shipped. Runs in Squad View terminals with constellation name identity. |
| Umbra | Background research. Scans external sources, stages findings, reports to Hoshi. | Claude Haiku | Shipped. Full pipeline: edge function runs, Hoshi auto-reviews, promotion to memory bank. |
| Kosmos | Memory operations. Grooms, labels, deduplicates, and heals the knowledge base. | Edge functions | Shipped. Three edge functions (mine, heal, groom) on pg_cron schedules. |
| Atlas | Documentation agent. Generates and maintains this docs site from codebase analysis. | Claude Sonnet | Shipped. Agent script, GitHub Actions workflow, and session-wrap dispatch hook all operational. |
Active Focus Areas
Atlas documentation system is now operational -- the agent reads source material (CLAUDE.md, MEMORY.md, session logs) and keeps these docs synchronized with the codebase through daily runs and session-triggered updates.
Near-Term Priorities
These are the highest-priority items in the backlog:
Atlas UI integration -- Adding Atlas to the Sidespace dashboard with an agent card and dynamic cost tracking. Currently the agent runs externally but isn't visible in the UI.
Design consistency and usability -- Visual design audit across views, adding note titles (currently untitled), and card pinning for the Hub dashboard.
Performance optimization -- Hoshi response latency tuning for the Gemini engine, terminal rendering improvements, and general polish passes.
Operational hygiene -- API key rotation, dependency updates, and architecture documentation formalization.
Open Questions
Some directions are clear in intent but uncertain in implementation:
Hoshi-to-Squad coordination. Hoshi can chat alongside Squad View terminals, but it cannot yet see what is happening inside them. Giving Hoshi awareness of running agents and their output is a near-term goal. Full AI orchestration -- Hoshi dispatching work to agents autonomously -- is deliberately deferred until the manual workflow is proven.
MCP as app control surface. The MCP server currently handles data operations (tasks, notes, memories). Extending it to control Sidespace's own surfaces -- creating terminal tabs, opening preview browsers, running headless tests -- would let agents interact with the app programmatically. The architecture supports it; the tool definitions are not yet written.
Memory system evolution. The knowledge bank has 600+ embedded memories with hybrid search (BM25 keyword backbone, pgvector semantic boost). The open question is how aggressively to automate memory lifecycle -- when Kosmos should merge, archive, or deprecate memories without human review.
Multi-user support. Sidespace is currently single-user. The data model (Supabase with RLS) can support multiple users, but the product assumptions (one Hub, one set of project tabs, one Squad View) would need rethinking.