onto

Onto

A Claude Code plugin that performs multi-perspective verification of logical systems using an agent panel (verification agents + Philosopher) and automatically builds ontologies from analysis targets.

Designed with inspiration from ontology structures, applicable across domains regardless of field -- software, law, accounting, and more.

Two core capabilities:

• Verification (review): The agent panel independently and in parallel performs multi-perspective verification on scope-defined targets
• Build: Incrementally constructs ontologies from scope-undefined analysis targets (code, spreadsheets, databases, documents) using integral exploration

Installation

Run the following commands in order within Claude Code:

/plugin marketplace add kangminlee-maker/onto
/plugin install onto@kangminlee-maker/onto

Update

/plugin install onto@kangminlee-maker/onto

Or if installed via git clone:

cd ~/.claude/plugins/onto && git pull

When upgrading from a previous version, run the migration since the global data path has changed:

./migrate-sessions.sh

If the learning storage structure has changed (3-path to 2-path + axis tag):

./migrate-learnings.sh

Migration targets:

• ~/.claude/agent-memory/ to ~/.onto/ (global learnings and domain documents)
• .claude/sessions/ to .onto/ (project session data)
• domain: + secondary_domains: → domains: unordered set in .onto/config.yml

Domain Document Installation (Optional)

Installing the domain base documents included with the plugin applies domain-specific expert criteria during reviews.

# Interactive -- select domains
./setup-domains.sh

# Install all
./setup-domains.sh --all

# Install specific domains only
./setup-domains.sh software-engineering finance

Domain	Description
software-engineering	Code quality, architecture, type safety, testing strategy
llm-native-development	LLM-friendly file/folder structure, ontology-as-code
finance	Financial statements, XBRL, accounting equation, cross-company comparison
business	Business strategy, revenue recognition, ROI, change management
market-intelligence	Market analysis, competitive intelligence, risk assessment, data reliability
accounting	K-IFRS, double-entry bookkeeping, tax adjustments, auditing
ontology	Ontology design, OWL/RDFS/SKOS, classification consistency
ui-design	UI layout, interaction patterns, accessibility, responsive design
visual-design	Visual hierarchy, color systems, typography, brand consistency

Usable without domain documents (verified using general principles). Domain document installation is also suggested when running /onto:onboard.

Getting Started

/onto:onboard                        # Set up project environment
/onto:review {target}                # Run agent panel review (interactive domain selection)
/onto:review {target} @ontology      # Run with specific domain
/onto:review {target} @-             # Run without domain rules
/onto:review {target} --codex        # Run in Codex mode (Claude tokens ~80% reduced)
/onto:ask-logic {question}           # Ask an individual agent
/onto:build {path|GitHub URL}        # Build ontology from analysis target

Domain Selection

Each process execution selects a single session domain. Three ways to specify:

Method	Syntax	Behavior
Explicit	@{domain}	Non-interactive, uses specified domain
No-domain	@-	Non-interactive, no domain rules applied
Interactive	(omit)	Analyzes target, suggests domain, user confirms

Project domains and execution mode are declared in .onto/config.yml:

domains:
  - software-engineering
  - ontology
output_language: ko
execution_mode: codex          # agent-teams (default) | codex
codex:
  model: gpt-5.4               # omit → ~/.codex/config.toml
  effort: xhigh                 # omit → ~/.codex/config.toml

domains: is an unordered set — order does not matter, no domain has priority over another. Domains can also be selected per session without pre-declaring them.

execution_mode: codex delegates reviewer passes to OpenAI Codex, reducing Claude token usage by ~80%. Tradeoff: deliberation (agent-to-agent exchange) is not possible. Currently supported for review only. Append --codex or --claude to override per-command.

Agent Configuration