Skip to content

weyucou/ibuki

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ibuki

A multi-tenant, multi-project, version-controlled specification and feature management system for AI agent development — so that human-authored intent and AI-executed implementation stay in sync.

Goals

Mission

Provide AI agents (and humans) with a single, queryable source of truth for what a project is building, and a structured workflow for proposing, evaluating, locking, and releasing feature changes.

Architecture

Dimension Decision
Deployment Hosted MCP service (single instance, multi-tenant)
Tenant boundary One tenant per GitHub org (e.g. weyucou, monkut)
Interface MCP server (Claude-native). REST may follow in v2
Auth ibuki-issued tokens, per tenant (decoupled from GitHub OAuth)
Feature granularity One spec per GitHub issue / PR

Primary outcomes

  1. Single org-wide source of truth — every project under a tenant has its feature specs in ibuki, queryable over MCP.
  2. Role-chunked agent consumption — agents fetch role-scoped slices (plan / develop / review) rather than full-document dumps.
  3. Bi-directional code ↔ spec evaluation — agents call ibuki check to ask either direction: "does this code match the locked spec?" or "does this proposed spec match the current code?"
  4. First-class git-style artifacts — locked specs are versioned, diffable, retirable; release manifests are immutable once cut.

Core workflow — proposal-driven

Submit  →  Rule check  →  LLM judge  →  Verdict  ↔  Proposer response  →  Lock  →  Release cut
                                       (iterative loop until convergence)
  1. Submit — a human or remote agent issues a change request (new feature or modification) via MCP.
  2. Rule check — deterministic engine validates schema + uniqueness + conflict detection (no duplicate IDs, no overlapping proposals on the same component).
  3. LLM judge — on rule pass, an LLM evaluates intent, scope, and consistency with existing locked specs. Verdict: accept, reject, or counter-propose (with revised draft + reasoning).
  4. Iterative response — proposer can accept the verdict, reject it (escalates), or counter-propose back. Loop continues until convergence.
  5. Lock — accepted proposals become immutable and queued for the next release.
  6. Release cut — a human or designated agent manually cuts a release. Locked-since-last-cut proposals bundle into a versioned immutable manifest. Versioning scheme is per-tenant configurable (semver / date / sequential).

Non-goals

  1. Not replacing GitHub issues — issues describe the next change; ibuki specs describe what is being built.
  2. Claude-first, not LLM-agnostic — initial target is Claude-family agents via MCP. Cross-vendor portability is a v2 concern.

Feature Pillars

# Pillar Description
1 Spec schema Per-feature schema (per issue / PR); rule-validated
2 MCP server + tenant auth Hosted multi-tenant MCP interface; ibuki-issued tokens; per-org routing
3 Proposal workflow Submit → rules → LLM judge → accept/reject/counter → lock; iterative response
4 Release management Manual release cuts; per-tenant versioning; immutable release manifests
5 Conformance check ibuki check MCP tool — code↔spec bi-directional evaluation, on agent request
6 Agent consumption Role-chunked spec output (plan/develop/review slices) over MCP

The current issues that implement these pillars are tracked on the Ibuki project board.

Per-pillar design docs land under docs/pillars/ as each pillar's design is locked:

Schema

The feature change-request schema (pillar 1) is the atomic unit of the ibuki workflow.

Validate a payload locally:

uvx check-jsonschema --schemafile schema/v0.1.0/proposal.schema.json <your-proposal>.json

Related Repositories

Repository Purpose
weyucou/ibuki Meta/spec repo (this repo) — schema, design docs, project board
weyucou/ibuki-backend Django 5.2 LTS backend + MCP server
weyucou/ibuki-infra AWS CloudFormation infrastructure (Postgres + Fargate)

Status

Pre-implementation. Goals are locked as of 2026-05-18 (see #6). Pillar issues are being re-scoped against the locked vision before implementation work begins.

About

Spec management system for AI agent development — canonical schema, lifecycle, versioning, and drift detection

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors