Skip to content

teaglebuilt/software-factory

v0.1.0

Repository-scoped skills for architect, developer, reviewer, tester, security and release roles

GitHub

Repo-scoped Software Factory for kagent 1.x

A reusable Helm application chart that installs one AgentTemplate + one Agent for each enabled role, using existing Harness, ModelConfig and MCP resources in namespace ai. No additional kagent controller, GPU deployment, Postgres, Agent Substrate, or Helm CRDs are installed.

Recommendation: install the chart per repo; share the platform

mlops / namespace ai
  shared: kagent controller + Harness(kagent) + ModelConfigs + Substrate + MCP
  release factory-showyourdream: architect / developer / reviewer / tester / security
  release factory-homelab:       architect / developer / reviewer / tester / release
  release factory-anotherrepo:   architect / developer / reviewer / tester

Each release has unique named agents (factory-showyourdream-architect, etc.) with a repository-scoped prompt and selected tools. Helm release scoping is not a security boundary. Enforce repo scoping in GitHub App tokens and MCP.

Prerequisites

  • kagent v1alpha3 CRDs/controller running in target namespace (your ai).
  • Existing Harness/kagent and ModelConfig/llmd-vllm-config in that namespace.
  • RemoteMCPServer/kb-retrieval in ai, or change architect tool bindings.
  • For working code edits: an independently secured RemoteMCPServer exposing repo-scoped Git read/write operations, with discovery-verified names. Your existing kind: MCPServer GitHub resource is NOT directly usable as RemoteMCPServer; point a RemoteMCPServer at a genuine HTTP MCP endpoint.
  • For executing tests: CI/test-runner MCP or a controlled sandbox.

Default role setup is safe but intentionally limited: architect can retrieve knowledge; developer cannot mutate code until you connect a Git tool server. Do not claim the factory is end-to-end autonomous until those dependencies exist.

Install

# From directory containing this README
helm lint charts/software-factory -f examples/values-showyourdream.yaml
helm upgrade --install factory-showyourdream charts/software-factory \
  -n ai -f examples/values-showyourdream.yaml
helm upgrade --install factory-homelab charts/software-factory \
  -n ai -f examples/values-homelab.yaml

kubectl get agenttemplate,agent -n ai -l app.kubernetes.io/part-of=software-factory
kubectl get harness,modelconfig,remotemcpserver -n ai

Before deployment, render and review with helm template. To uninstall only a repo's agents: helm uninstall factory-showyourdream -n ai (this does NOT remove the shared controller, Harness or MCP infrastructure).

Enable Git-backed skills

  1. Copy plugin/software-factory/ to plugins/software-factory/ inside your teaglebuilt/agent-plugins repository or another Git repository.
  2. Commit and capture the full 40-character SHA.
  3. Set plugin.enabled=true, plugin.url, plugin.commit, plugin.path.
  4. Reinstall/upgrade. Source path is under source.path, outside source.git.

The bundled plugin has six distinct skills: architecture-design, implement-change, review-change, test-change, security-assessment, release-plan. Each role only loads its own skill. For the default role prompts, the plugin is optional and is not needed to render or start the chart.

Role / harness choices

  • kagent is the default for all roles; works with local llmd-vllm-config. This is intentionally a model-agnostic chart, not a second inference server.
  • Current Qwen/Qwen2.5-3B-Instruct may have poor multi-step reliability. Use a stronger local model when feasible, and keep tests/PR gates external.
  • You can select a different Harness and ModelConfig per role, but validate compatibility. Claude harness does not honor partial MCP tool allowlists.

A2A or not?

A2A is always present for Agents at your kagent controller. Agents are invoked through /agents/{namespace}/{name}. Do not install a separate A2A deployment. For external orchestrators or coding clients, kagent also offers MCP access to agent sessions. Use an external durable coordinator for an actual workflow.

router.enabled=true creates an optional conversational dispatcher using subAgent bindings. It hands over the conversation (one Actor) and is NOT an architect/developer/tester/reviewer pipeline. See docs/orchestration.md.

GitOps / ArgoCD

The optional examples/argocd-applicationset.yaml defines an ApplicationSet for homelab and showyourdream, with one chart release each. Move the chart into kubernetes/charts/software-factory before applying; the path in the example is proposed, not present in your current Git repository. Per-repo tool bindings are still required for developer/tester access.

Repo-specific bindings

roles.<role>.toolBindings is an array. Every entry needs:

- serverName: repo-github-write
  tools: [create_branch, push_files, create_pull_request]
  requireApproval: true

Optional serverNamespace references a RemoteMCPServer outside ai, but the server must explicitly allow access from ai, and the controller must watch both namespaces. Tool names above are EXAMPLES only; discover the real names from status.discoveredTools before use.

Testing

Agent smoke-test prompts and pass/fail rules are in tests/evaluation-prompts.md.

bash scripts/validate.sh
helm template factory-showyourdream charts/software-factory -n ai \
  -f examples/values-showyourdream.yaml
kubectl apply --dry-run=server -f /path/to/rendered.yaml

This package was authored against kagent 1.x api.kagent.dev/v1alpha3. The chart does not mutate your connected GitHub repo or live Kubernetes cluster.