Problems it solves
Teams already use coding agents. What they lack is a repeatable operating model. Read the six failures, then how the architecture answers them.
Context explosion
Every skill, brief, and wiki page under the IDE folder is auto-discovered. Cost and quality drop as the pack grows.
Glued to one IDE
Process mixed with editor wiring. Moving from Cursor to Claude Code meant copying and rewriting the tree.
Customer leakage
Ports, tracker projects, and folder names baked into skills. The next engagement could not reuse the pack.
No productized install
Copying last year’s .cursor folder does not scale. Architects need project, user, and later org scopes.
One ladder for every ask
A “how does tax work?” question should not start PM → BA → developer. Workflows are first-class.
Hard to add a process
A new architecture review should be a workflow JSON + skill + config chain — not a new platform.
How it is wired
A thin IDE adapter calls the receptionist. The pack owns process. The loader allowlists files for the current step only.
Capabilities
Core delivery is installed with init. Knowledge, plugins, and agent-run observability are opt-in.
Workflows
ask, feature-development, Jira story/epic/bug, QA bootstrap.
Knowledge base
Opt-in QA overlay from a Graphify graph, then human promote.
Plugins
Optional Graphify and Archify. The kit never vendors them.
Observability
Agent-run traces and scores to a local ledger, then Langfuse.
Planning gates
Human sign-off on requirements, architecture, and BA specs.
Allowlist loader
Each specialist reads only the files for the current step.
Feature flags
Toggle test design, Playwright, telemetry, Jira, and obs.
Agent wiki
Trigger-routed operational memory. Load one page, not the folder.
Who this guide is for
Same handbook, three entry points — pick by what you need to do.