ruc-datalab/EvoOntology

What it claims

The summary names the build-use-evolve-evaluate-publish loop but leaves out the concrete surface an engineer actually touches: the plugin install path, the slash commands, the runtime tool names, and how patches are scoped. This document carries those. The core framing the page adds: the Ontology Layer is treated as trainable agent state, not model weights — a builder initializes grounded semantic objects from the workload, and a separate evolution agent proposes bounded updates from interaction history, validating each Candidate against its Parent.

How it works

Two distinct agents, not one. The builder derives candidate concepts from the workload, verifies them against raw sources, and publishes ontology_v0. The evolution agent diagnoses recurring behavior in recorded trajectories, attributes each issue to one of three targets — Content, Tool, or Schema — and produces a localized patch rather than a full rewrite. Evaluation compares Parent and Candidate under identical data, agent, decoding settings, and interaction budget before promotion to ontology_vN+1.

The ontology is structured in three layers: a Content Layer (concepts, mappings, constraints, evidence, relationships) plus representation-rule and runtime-access layers. At runtime the Data Agent calls two tools — browse_semantics and resolve_semantics — with no further ontology configuration once built.

How to use it

Installed from the GitHub marketplace; no clone, venv, or pip step.

Claude Code:

claude plugin marketplace add MeiduoChong/EvoOntology
claude plugin install evoontology@evoontology
claude plugin list

Then in a new session: /evo-build, /evo-evolve, /evo-visualize.

Codex:

codex plugin marketplace add MeiduoChong/EvoOntology
codex plugin add evoontology-codex@evoontology
codex plugin list

Then: $build-ontology, $evolve-ontology, $explore-ontology.

For benchmark integration, each environment implements an EvolutionAdapter preserving its native rollout/evaluation protocol; list environments with python -m benchmarks list.

Evidence

The page states directionally that the builder-constructed Initial Ontology Layer beats ReAct-without-ontology, and self-evolution adds a further gain on all three benchmarks (DDR-Bench, InsightBench, BIRD), across the four-backbone analysis subset. Note the caveat: the summary's benchmark numbers come from a four-backbone analysis subset; DDR-Bench values are direct from Tables 2/8, but InsightBench and BIRD figures are one-decimal means read off Figure 3, not primary reported numbers. Full six-backbone results and protocols are in Tables 1, 3, 4 — the page itself does not print those totals.

Limits

Data-source support is uneven: SQLite has built-in read-only task replay, but other data sources require host tools with explicit observation recording — evolution quality depends on trajectories you must instrument yourself. Paper is arXiv 2609.15779 (2026); no license is stated on the page, so verify before building on it.