OptiSkill

OptiSkill is an agentic skill that enables agent learn how to analyze, and manipulate Pyomo optimization models using natural language.

Unified Skill

The SKILL.md file contains six analysis workflows in one self-contained skill:

Workflow	Mode	Description
A	Model Description	Generate narrative overview + structured component reference
B	Retrieval	Read current values (params, vars, constraints, objective)
C	Sensitivity	Dual values / shadow prices for marginal impact analysis
D	What-If	Modify a parameter, re-solve, compare outcomes
E	Why-Not	Force a decision variable, explain cost penalty
F	Code Execution	Custom Python code for complex model manipulations

Priliminary results

Overall results by test suite.

	Questions	OptiChat	OptiSkill
Quantitative verification	96	96/96 (100.0%)	70/96 (72.9%)
Qualitative evaluation	37	30/37 (81.1%)	8/37 (21.6%)
Total	133	126/133 (94.7%)	78/133 (58.6%)

Per-query-type accuracy.

Query Type	Evaluation	Questions	OptiChat	OptiSkill
Retrieval	Quantitative	39	39/39 (100.0%)	32/39 (82.1%)
Sensitivity	Quantitative	18	18/18 (100.0%)	5/18 (27.8%)
What-if	Quantitative	39	39/39 (100.0%)	33/39 (84.6%)
What-if	Qualitative	11	9/11 (81.8%)	3/11 (27.3%)
Why-not	Qualitative	26	21/26 (80.8%)	5/26 (19.2%)
All		133	126/133 (94.7%)	78/133 (58.6%)

How to Run

1. Read SKILL.md.
2. Follow Prerequisites (0a–0c) to install deps and build the demo model.
3. Classify the user's question using the Query Classification tree.
4. Execute the matching Workflow (A–F).

Dependencies

All pip-installable, no commercial licenses:

• pyomo — algebraic modelling language
• cloudpickle — model serialisation
• highspy — open-source LP/MILP/QP solver (HiGHS); glpk as fallback

Demo Model

The Stigler Diet Problem is bundled inline: 19 food items, 9 nutrients, cost minimisation.