The product
Torii packages a SUMO expert skill with local MCP tools, so Codex can reason about intent and run bounded checks on real SUMO artifacts.
Torii
Task-Oriented Road Infrastructure Intelligence
One prompt, structured SUMO work.
Agent plugin for SUMO: one-sentence workflow routing plus local MCP tools for OSM-to-SUMO work.
Use Torii to clean the Ingolstadt city-center network from OSM and compare it with TUM-VT/sumo_ingolstadt.
Intent
Observe
Build
Audit
Diagnose
Evidence
What Torii changes
It makes the agent ask what result the user needs, observe network and metric feedback, and then choose the smallest defensible next action.
Who should use it
Researchers and builders who need OSM-to-SUMO import, TLS review, routeability checks, controller workflow guidance, and evidence-bounded SUMO repair.
It treats warnings, bad metrics, disconnected routes, TLS mismatches, and map disagreements as feedback about the model.
Torii uses a closed loop: user intent, observed SUMO artifacts, feedback diagnosis, smallest model or code change, rerun, and evidence-bounded report.
- OSM extracts need tiling, retry, deduplication, and road-class filtering.
- Generated SUMO TLS candidates need physical-intersection review.
- Reality baselines are region-aware: use Google Maps where appropriate, and Amap/Gaode, Baidu Maps, Tencent Maps, inventories, signal plans, or field photos for mainland China.
- Routeability probes expose disconnected roads before result claims.
- Controller work must be planned, tested, and interpreted against completion evidence.
- Bad metrics are feedback for diagnosis, not the final optimization target.
Current scope
Current tools cover bounded environment checks, config preflight, smoke runs, evidence bundles, OSM network construction, area confirmation, multi-source TLS review tables, passenger connectivity checks, routeability probes, and Netedit launch evidence.
Unconfirmed place-name-to-bbox geocoding, full city-scale intelligent cleanup, max-pressure controller generation, and controller-log inspection are roadmap tools, not finished promises.
It connects reasoning to bounded execution.
The workflow router classifies the request first. The skill interprets feedback, and the MCP server runs bounded checks that return structured observations.
Ongoing project
Workflow screen -> target, state, deviation, next-step plan.
New experiment
Experiment intake -> readiness record -> SUMO Experiment Plan.
Code change
Develop and verify -> RED/GREEN/REFACTOR record -> completion evidence.
Failed or suspicious run
Debugging helper -> symptom evidence -> fix, rerun, or claim demotion.
Results and claims
Evaluate and report -> outputs, metrics, baselines, validation, claim boundary.
simulation-helper-skill-for-eclipse-sumo
Use for OSM modeling, TLS audit, controller workflow planning, metric feedback diagnosis, and evidence-bounded SUMO claims.
debugging-helper-skill-for-eclipse-sumo
Use for closed-loop diagnosis of route, TraCI, TLS, demand, detector, output, seed, completion, and reproducibility failures.
Design approach
The skills combine official SUMO documentation, SUMO community troubleshooting patterns, public traffic-simulation code lessons, Socratic intake, confirmed experiment planning, TDD for SUMO/TraCI code changes, evidence-before-completion review gates, and privacy-safe field lesson capture.
Use scenarios
Start from the user's situation, then load only the references needed for that path.
- Ongoing project: route-project-workflow -> next-step plan.
- New experiment: interactive-experiment-intake -> readiness route -> SUMO Experiment Plan.
- Code change: develop-and-verify-code -> implementation -> completion evidence.
- Failed run: debugging helper -> closed-loop diagnosis -> fix or demotion.
- Results or claims: evaluate-and-report-results -> evidence class -> claim boundary.
Copy-ready prompts for real projects
Start with one sentence. Torii should choose the workflow, ask only blocking questions, and report evidence before claims.
Use Torii to clean the Ingolstadt city-center network from OSM and compare it with TUM-VT/sumo_ingolstadt.
Audit this TraCI signal controller before I compare it with the fixed-time baseline.
This SUMO run finishes, but tripinfo and summary disagree. Use the debugging helper and tell me whether the run is valid.
Controller A has lower travel time, but some vehicles are unfinished. What can I claim and what must be demoted?
The skill missed a failure mode that I later fixed. Turn my solution path into a privacy-safe field lesson candidate.
Reference modules
This SUMO skill is split into focused Markdown modules so the agent can load the right evidence rules only when needed.
Torii reasoning references
- route-project-workflow.md
- preflight-sumo-environment.md
- interactive-experiment-intake.md
- plan-experiment.md
- develop-and-verify-code.md
- model-osm-detectors.md
- learn-sumo-knowledge.md
- audit-sumo-controllers.md
- compare-corridor-perturbations.md
- evaluate-and-report-results.md
- capture-field-lesson.md
- release-project.md
Debugging audit references
Audit checks
- Network, route, config, detector, and additional-file consistency.
- TLS phase, movement-green, yellow, all-red, and NEMA evidence checks.
- Paired controller comparisons across seeds, demand, horizons, and outputs.
- Completion-aware reporting for stopped, unfinished, teleported, or invalid runs.
- Metric definitions for travel time, delay, stops, throughput, queue, emissions, and fairness.
- Claim boundaries for baselines, ablations, sensitivity runs, and debugging fixes.
- Field lesson capture for user-discovered fixes that should become reusable skill guidance.
- TDD records for SUMO/TraCI code changes before implementation claims.
- Fresh verification evidence and review findings before completion claims.
- Confirmed SUMO Experiment Plan after intake and before execution.
Examples
The repository includes a unified signal-control audit example and an Ingolstadt OSM-to-SUMO network comparison against a manually cleaned TUM reference.
Install Torii
Install the Codex plugin to get both the reasoning skill and local MCP tools.
codex plugin marketplace add Tarard/Torii-SUMO --ref main
codex plugin add torii-sumo@torii-sumoLimitations
This repository provides agent instructions, checklists, and audit procedures. It does not certify that a SUMO experiment is correct, publishable, or officially validated.
Eclipse SUMO is a trademark of the Eclipse Foundation. This project is independent and is not affiliated with, endorsed by, sponsored by, or maintained by the Eclipse Foundation, the Eclipse SUMO project, or DLR.