HiveMind is the control plane behind how I run multiple AI agents without losing context between sessions. Instead of copy-pasting chat logs, every task, handoff, and decision lands in an Obsidian vault that Codex, Claude, Grok, and custom workers can read back on the next run.
The monorepo ships four apps and a dozen packages: a REST API over the vault filesystem, a BullMQ orchestrator backed by Redis, a password-gated Next.js dashboard with live SSE updates, Slack/Discord bot listeners, and per-agent workers that poll their own scratchpads. Phases 1 through 8 are complete — routing, handoff execution, Docker Compose, parallel swarm runs, Grok streaming, vault search, synthesis, and scheduled pipeline jobs.
The architecture is deliberately boring in the right places. Agents register with runtime modes (api, cli, stub). Handoffs carry conflict resolution (claimed_by, stale timeout). The promoter package moves finished notes from Inbox into Projects or Knowledge Base. TraderBot, this portfolio, and game sprints all write outputs into vault/02-Projects/ so downstream agents inherit full context.
I built HiveMind because agent wrappers collapse without durable memory. The bet is that a filesystem vault + queue + dashboard is enough infrastructure for a solo operator to run a micro-agency, trading signals, and three game repos without hiring a platform team.
