Spine — the governed delivery loop
A request flows top to bottom — through comprehension and planning, into an execution loop that pauses at two human gates — and out as a reviewed pull request.
You ask for something — from a terminal, your AI assistant, or the web inbox. Same engine behind all three.
Command line
55 commands · the main surface
Assistant plugin
Claude Code & Codex call it as tools
plugin/Web app + API
the operator inbox — approvals live here
registry/Before writing anything, Spine reads. Your requirement becomes structured intents; your code and your docs become a graph.
Read the requirement
Confluence · Jira · OpenSpec · files → intents
intake/Size up the repo
languages, framework, what kind of job
catalog/Write it down
a committed, code-true knowledge base
knowledge/Product Knowledge Graph
the source of truth · pkg/Every fact points at a file and a line. 8 node kinds · 11 edge kinds, built from code and docs. No LLM involved — so the same commit always gives the same graph. Everything above and below reads from it.
The work becomes a typed, validated plan — which tools, which budget, where the approvals sit. Not free-form prompting.
Choose the approach
objective → agents + tools
planner/The plan, typed
validated before anything runs
ir/ · GraphIRReusable parts
versioned templates + tool contracts
registry/Stop. A human approves before a single line is generated — nothing has been written yet.
Now it runs. Each step is one contract-checked call, verified against schema, evidence and policy. Tools get credentials only at the moment of use.
Run one step
perceive → plan → act → observe
runtime/Check every step
schema · evidence · policy
runtime/Hand out credentials
only at the moment of use
gateway/Survive the pause
resumable across gates and restarts
temporal/Code is generated grounded in what already exists, tested, and retried until the tests pass. Then it's reviewed.
Build the change
spec → code → tests → PR
sdlc/Write, test, retry
until the tests are green
agentic/Review it
secrets · security · style
codereview/Same engine, other jobs
PR reviewer, codebase auditor
personas/Stop again. Until this moment nothing has been pushed to your repo or your tracker.
The pipeline ends here — deployment stays yours.
Every step above was traced and recorded — an append-only trail you can replay.
Used across every layer.
Model access
one client, any provider
core/Use outside tools
databases, Atlassian — governed
mcp/Keep artifacts
runs, bundles, reports
storage/Tell someone
Slack on gates
notify/Measure it
the eval harness
evals/Domain meaning
optional — off unless configured
spine/