GitHub Projects as the Agent-Native Support Control Plane

·https://github.com/skillrecordings/support
repoaisupportgithubwebhooksagent-loopsslackfrontinngesttypescript

GitHub Projects plus webhooks gives support agents a simple shared control plane that maps cleanly onto joelclaw-style event feedback loops.

Skill Recordings Support is an agent-first customer support platform for Skill Recordings products like Total TypeScript and Pro Tailwind. Front stays the conversation source of truth, while an AI agent drafts replies and humans approve them through Slack or a Next.js dashboard.

The clever bit is the operating surface around the work. Joel noted that the repo is being built primarily with clawdbot, now moltbot, syncing into a GitHub Projects board and feeding GitHub webhooks back into the bot. Linear was considered, but GitHub CLI plus webhooks is boring in the best way: easy for agents to read, easy for humans to groom, and already close to the code.

The product architecture is also a clean HITL support loop: Front webhook → Inngest workflow → classify → draft → Slack approval → send through the Front API. High-trust, high-confidence conversations can auto-send. That is the useful tension: support work becomes a queue of reviewable decisions, not a pile of vibes and manually copied context.

The stack is practical agent scaffolding: Bun workspaces, Turborepo, PlanetScale with Drizzle, Upstash Vector and Upstash Redis, Claude through the AI SDK, plus Axiom and Langfuse for traces. Not exotic. Just enough machinery to keep the loop observable and auditable.

Key Ideas

  • Front remains the source of truth for customer conversations, which keeps the support record out of random bot state.
  • Slack approvals make the human-in-the-loop step lightweight instead of forcing reviewers into a custom app for every decision.
  • GitHub Projects works as an agent-readable support control plane because GitHub CLI and GitHub webhooks are already native agent tools.
  • Inngest turns inbound support messages and approvals into durable workflows instead of one-off scripts that forget where they are.
  • Trust scoring creates a path from manual approval to auto-send when confidence and history are strong enough.
  • The private skill CLI gives the team a local operator surface for auth, diagnostics, and setup, including skill doctor and skill auth oauth-spike.

Links

← All discoveries