How it runs, and what keeps it honest
rules
hooks
an ideaa transcriptlegacy docsa repo
→
spec-builder
one contract→
Agent(s)
→
docs + graphplancode + graph
state: prompt, run history, task board
Tailored, not comprehensive
Most kits install everything, then hope the project needs it.
every agent, a hundred skills
you pay context for all of it
seats no module owns
advice too general to act on
→
7 to 15 of the 16 seats
derived from the contract
and the modules that exist
one dev agent per module
Skill Discovery
The problem: you cannot choose from a catalog you have never read.
requirements.txtGemfile*.csprojbuild.gradle
→
detected tech stack
→
skills that matchyou choose, or not
Skill Wire
The problem: an installed skill nobody is told to use is dead weight.
a chosen skill
/skill-wire→
the agent that needs it
harness-view
The problem: how do you know it is actually wired right?
Flow / Graph
every wire, read off disk
Assess
a score, and what is wrong
Toggle
switch any rule or hook off
Watch
re-scan as files change
Chaos in.
Controlled, inspectable work out.
AGENT HARNESS BOOTSTRAP