KBase v0.2.1 — Deliverable modifiers and plan mode

Introduction

KBase is a structured, self-organizing knowledge base for AI coding agents. It persists agent knowledge — rules, references, decisions, skills — across sessions, so agents start informed instead of cold. Every directory follows a consistent AGENTS.md + INDEX.md pattern, with a tiered authority model (ROOT > rules > refs > notes > external) and a communication protocol for human–agent interaction.

Usage

Point your AI coding agent's config to the KBase/ directory. The agent reads AGENTS.md on session start and self-maintains the knowledge base.

AGENTS.md → ROOT/AGENTS.md → ROOT/communication.md → ROOT/instructions.md → rules/INDEX.md → skills/INDEX.md → notes/INDEX.md

Highlights

• This release introduces deliverable modifiers — suffixes on [ ... ] and [[ ... ]] that control how artifacts are delivered. Three modifier families: read-only (RO), write-allow (+W), and skill-first (SK).
• A new Plan Mode subsection formalizes when agents should ask before proceeding with major edits.

What's Changed

Features

• Add (RO) read-only deliverable modifiers — analysis/review only, no product code changes (#6)
• Add (+W) write-allow deliverable modifiers — explicit permission to edit product code (#6)
• Add (SK) skill-first deliverable modifiers — explicitly search skills and rules before proceeding (#6)

Improvements

• Add plan-mode bias principle to ROOT (#6)
• Extend compaction recovery chain to include rules/INDEX.md and skills/INDEX.md (#6)

Documentation

• Update protocol summary tables in all 5 README translations (#6)

Compatibility

• Agent: Any AI coding agent that reads AGENTS.md
• Platform: Any

KBase v0.2.0 — Meta Rules & Meta Skills

What's New

Meta Rules (rules/INDEX.md)

Six auto-triggering meta rules that govern agent behavior during conversations — no user command required:

• auto-save — Persist findings and user decisions to notes/ immediately
• recall-first — Search KBase before any external lookup
• deduplication — One concept, one file; extend existing entries instead of duplicating
• promote-audit — Verify via external sources before promoting notes/ → refs/
• staleness-review — Check refs/ entry age; re-verify or demote when stale
• skill-extraction — Auto-extract repeated workflows into skills on second occurrence

Meta Skills (skills/INDEX.md)

Two skills that operate on the skill system itself:

• write-skill — Standardized workflow for authoring new skills
• audit-skills — Read-only integrity check across the skill system

Fixes

• Bootstrap chain — Added missing ROOT/communication.md to recovery chain in ROOT/compaction.md and bootstrap section in all 5 READMEs
• Compaction count — Corrected "first three" → "first four" in ROOT/AGENTS.md

KBase v0.1.1 — Persistent Knowledge Base for AI Coding Agents

KBase

Structured, self-organizing knowledge base that persists agent knowledge across sessions.

What is KBase

KBase stores rules, references, decisions, and skills so AI coding agents start informed instead of cold. Knowledge accumulates across sessions — agents read AGENTS.md on start and self-maintain the base.

Structure

Directory	Role
ROOT/	Meta-principles, bootstrap, commands
rules/	Constraints and conventions
refs/	Verified external knowledge
notes/	Project insights and decisions
skills/	Automation workflows
plans/	Auto-generated task plans
env/	Environment configurations
hooks/	Event-driven callbacks
perms/	Secrets and credentials (gitignored)

What's New in v0.1.1

Review-Fix Notation

Two new notation forms for the communication protocol:

Notation	Meaning
[ X ] : ( Y )	Soft — Review scope X against expectation Y; fix if unsatisfied. Best effort.
[[ X ]] : (( Y ))	Hard — Deep review X against Y. Must satisfy; violation = blocked.

The soft form composes existing [ ] (optional deliverable) with ( ) (prefer constraint). The hard form composes [[ ]] (must deliverable) with (( )) (must constraint), requiring thorough review and strict satisfaction.

Features

• Active persistence — agents save findings immediately without being asked
• Tiered authority — ROOT/ > rules/ > refs/ > notes/ > external; higher tier overrides lower
• Communication protocol — structured notation for human–agent interaction
• Lifecycle transitions — notes promote to refs (verified) or rules (recurring constraint)
• Agent-agnostic — works with any AI coding agent that supports AGENTS.md
• Built-in commands — research, review, tidy, recall, save, promote

Quick Start

1. Clone: git clone https://github.com/hayabusa-cloud/KBase.git
2. Point your AI coding agent's config to the KBase/ directory.
3. Work. The agent reads AGENTS.md on session start and self-maintains the knowledge base.

MIT License — ©2026 Hayabusa Cloud Co., Ltd.

KBase v0.1.0 — Persistent Knowledge Base for AI Coding Agents

KBase v0.1.0 — Persistent Knowledge Base for AI Coding Agents

Structured, self-organizing knowledge base that persists agent knowledge across sessions.

What is KBase

KBase stores rules, references, decisions, and skills so AI coding agents start informed instead of cold. Knowledge accumulates across sessions — agents read AGENTS.md on start and self-maintain the base.

Structure

Directory	Role
ROOT/	Meta-principles, bootstrap, commands
rules/	Constraints and conventions
refs/	Verified external knowledge
notes/	Project insights and decisions
skills/	Automation workflows
plans/	Auto-generated task plans
env/	Environment configurations
hooks/	Event-driven callbacks
perms/	Secrets and credentials (gitignored)

Features

• Active persistence — agents save findings immediately without being asked
• Tiered authority — ROOT/ > rules/ > refs/ > notes/ > external; higher tier overrides lower
• Communication protocol — structured notation for human–agent interaction ([[ ]], (( )), ( ) > ~( ), etc.)
• Lifecycle transitions — notes promote to refs (verified) or rules (recurring constraint)
• Agent-agnostic — works with any AI coding agent that supports AGENTS.md
• Built-in commands — research, review, tidy, recall, save, promote

Quick Start

1. Clone: git clone https://github.com/hayabusa-cloud/KBase.git
2. Point your AI coding agent's config to the KBase/ directory.
3. Work. The agent reads AGENTS.md on session start and self-maintains the knowledge base.

MIT License — ©2026 Hayabusa Cloud Co., Ltd.