NEXUS

NEXUS is the Controller of the OpenFlows control plane - a long-lived process running in the openflows-nexus Coder workspace. It syncs GitHub issues into typed tickets, assigns work to FORGE workers inside isolated Coder workspaces, and drives every agent through the gated lifecycle, so every action inherits Coder's identity, audit trail, and workspace governance.

Capabilities

• Autonomous GitHub Issue Discovery
• Multi-worker Task Assignment inside Coder workspaces
• Self-healing Pipeline Recovery (reconcile())
• CI Readiness Enforcement - prioritizes a CI-setup ticket first when a repo has no CI
• Human-in-the-loop Escalation (awaiting_human) via Slack, Discord & WhatsApp
• Live Control Surface - web control panel & CLI with a no-restart agent registry
• Multi-tenant Isolation - one Coder server, many teams via Coder RBAC + Redis keyspaces
• SharedStore State Supervision

How it works

1. 01 Issue Discovery

   Polls GitHub for open issues and syncs them into the SharedStore as typed tickets (T-001, T-002...) - the Redis store agents use to coordinate.

2. 02 Work Assignment

   Matches priority tickets to idle FORGE workers inside isolated Coder workspaces, respecting CI readiness and recovery state.

3. 03 Pipeline Supervision

   Runs a paced poll loop every ~15s across a declared flow graph, monitoring every phase across the Coder-backed agent fleet.

4. 04 Flow Recovery

   Runs reconcile() every cycle to detect and repair unmerged PRs, orphaned tickets, stale workers, and crashed workspaces or chats - bounded to 3 attempts, then escalates to awaiting_human.

Human Intervention

NEXUS is your command center. The team is autonomous in the happy path, but it is never unsupervised - you hold final authority at every meaningful decision point, and NEXUS is the single place where the whole team surfaces what it needs from you.

• Escalations surface here: NEXUS parks any ticket that reaches awaiting_human and notifies you via Slack, Discord, or WhatsApp - with the ticket ID, reason, workspace link, and GitHub link. It never lets an escalation deadlock the fleet or silently stall.
• Human gates: you approve or reject the testing and submit gates through the operator CLI (openflows gate decide) or the control panel. Your decision is recorded in the same versioned lifecycle record as every other transition, bound to a specific plan revision, review round, and commit head.
• Steer the fleet live: pause, drain, or target repos and issues, edit the live agent registry (roles, instances, model, skills, MCP) with no restart, and watch live status in one place.
• Full auditability: every agent action inherits a real Coder + GitHub identity, so you can audit who did what, when, and why.