SSOT

2.0.0

Architecture single source of truth as validated records instead of prose: bootstrap or repair a product, keep decisions/services/use-cases machine-checkable, and query them through a local index.

architecturedocumentationgovernancemcpssotverification

Install

Downloads and runs the installer. Cloning the repository is not required.

macOS, Linux and WSL

curl -fsSL https://agentkits.ominot.es/i/ssot | bash

Windows (PowerShell)

irm https://agentkits.ominot.es/i/ssot.ps1 | iex

Update an existing install

curl -fsSL https://agentkits.ominot.es/i/ssot | bash -s -- --update --auto

The update command recovers the token, endpoint and CLI list from the configs already on disk, so nothing has to be pasted again.

What the installer asks

Answered once and remembered, so reinstalling and updating stay silent. Pass --set id=value to answer up front, or --no-companions to decline the extras.

What is this repository?

mode=new New product Turn a PRD into the smallest confirmed architecture and its records
mode=existing Existing codebase default Derive records from the code that is actually there, then close the gaps

How should the agent read the SSOT?

retrieval=text Files only No server. The agent reads records directly. Simple, but costly on large repositories
retrieval=index Local index default Deterministic SQLite index: typed record queries, routing, validation, drift. No model, no network
retrieval=assisted Local index + model-assisted search Everything above, plus an opt-in sub-agent that narrows results. Spends tokens

Which model should assisted search use? (ignored unless retrieval is assisted)

model=deepseek-v4-flash opencode/deepseek-v4-flash default Fast and cheap; the default
model=claude-haiku anthropic/claude-haiku-4-5 Better narrowing, higher cost
model=claude-sonnet anthropic/claude-sonnet-5 Best narrowing, highest cost
model=env From SSOT_SEARCH_MODEL Any model string, taken from the environment at run time

Companion skills

Installed through their own ecosystem's CLI, after this kit is written. Declining one never fails the install, and an already-installed skill is left alone.

find-skills offered by default vercel-labs/skills Lets the agent discover current stack-specific skills instead of writing 2023 syntax from memory.

Contents

Skills

ssot

Commands

/ssot

Notes

The SSOT lives in docs/architecture as text records. The index in .agents/cache is derived and disposable; delete it any time and reindex.
Do not install this kit and the legacy howto-ssot kit into the same target: both register the /ssot command.
In assisted retrieval the server starts `opencode` as a child process to narrow search results. It spends tokens and needs opencode on PATH. Choose the index option if you do not want that.
Keep `ssot validate` in CI. Without it the record envelopes decay and the SSOT becomes a folder of markdown again.