English | 简体中文
Build an ontology layer from data and analytical goals, validate improvements, and explore concepts, evidence, and execution results. The plugin bundles the deterministic core and MCP, so users do not install a Python package in their projects.
| Entry point | Purpose |
|---|---|
$build-ontology |
Prepare questions, build the initial ontology layer, and open the result page |
$evolve-ontology |
Collect real tasks, run a missing baseline, validate improvements, and present the result |
$explore-ontology |
Explore questions, concepts, evidence, execution results, and version differences at any time |
You can also ask naturally: “Build an ontology layer for this data, focusing on sales analysis,” or “Improve it based on recent use.” Claude Code and Codex share the same Build, Evolve, and Explore workflow names; only invocation syntax follows each platform.
codex plugin marketplace add ruc-datalab/EvoOntology
codex plugin add evoontology-codex@evoontology
codex plugin listStart a new session after installation or an update so the new skills and MCP tools are loaded.
Questions expand coverage in this order: current requirements, history from the same data project, then evidence-backed exploratory questions. The host Agent extracts history from accessible context, project records, or user-specified sources; the plugin does not read the account's entire chat history automatically. Without history, the Agent inspects metadata and samples to propose grounded questions. Explicit scope takes precedence, and fixed benchmarks may use only their original construction partition.
prepare_ontology_workload preserves provenance, deduplicates questions, and selects by priority and topic coverage. start_ontology_task freezes a question and ontology version; resolve_ontology_task interprets concepts using that version. SQLite queries can run read-only through execute_ontology_query and automatically record real results. Other data environments use host-native tools, then record observed results with record_ontology_task_event and persist the trajectory with finish_ontology_task. ontology_workflow_status reports resumable in-progress tasks.
The plugin never fabricates execution results and cannot obtain complete trajectories from ordinary chat alone.
Build uses configure_ontology_project to create or reuse project settings. After semantic objects and question associations are saved, publish_ontology_build validates, activates, initializes state, and opens the result page.
Evolve supports quick, with two rounds, and full, with eight rounds by default. It reuses the project budget unless the user specifies one; both modes use the same evaluation criteria. Resuming a run keeps its frozen budget. Only trustworthy improvements are published; otherwise the current version remains active. finalize_evolution_run presents accepted or incomplete results. Presentation failure is reported separately and does not undo publication.
The result page includes a summary, known limitations, linked question and semantic subgraphs, real task outputs, evolution metrics, and public-task replays. Synthetic and real tasks must be reported separately. Successful execution does not imply correctness, and structural change does not imply quality improvement. Formal validation examples are not embedded in the result page, and in-progress validation metrics are not displayed.
visualize_ontology generates one offline HTML file at <workspace>/visualizations/ontology-layer-explorer.html and opens the browser once by default. In a headless environment, pass open_browser:false.
rolling_trajectory: production projects or cold starts; accumulate questions and trajectories from requirements and project use.fixed_split: benchmarks with fixed partitions. Construction supports Build and diagnosis, Validation is used only for the independent gate, and Held-out remains unseen before freezing. Registerconstruction_question_idsbefore importing questions.
Semantic MCP starts through .mcp.json. The default workspace is <project-root>/.evoontology/. Skill workflows are under skills/. The root evoontology/ package is the sole core source and is synchronized into the plugin with scripts/sync_plugin_core.py.
For publish_ontology_build, finalize_evolution_run, and visualize_ontology, pass presentation:"codex" and open_browser:false when the Codex browser panel is available. Open the returned presentation.browser_url, or browser_url for visualize_ontology, with the available open_in_codex tool using target:{type:"browser",url:browser_url} and placement:"right". Reuse an existing preview tab where possible.
This mode serves only the generated HTML on loopback and does not open an external browser. If the app tool is unavailable, return the working URL and file path without claiming that it opened. Explicit headless requests skip preview startup.