lagrangian-skills

v1.0.0 Complete Release Standard

This package is a complete v1.0.0 release. Completeness is defined by the presence of the core skill, reference docs, golden tests, eval scenarios, schemas, release notes, security policy, contribution policy, and a passing release validator.

Run:

python scripts/check_release.py
python evals/run_eval.py

Expected result: both commands pass with exit code 0.

A stable Agent Skill for constrained optimization using Augmented Lagrangian Methods (ALM), ADMM, KKT-style verification, infeasibility diagnosis, and capability-aware no-tool fallback. It is designed for agents that need to reason about budgets, capacities, safety limits, equality/inequality constraints, multi-objective trade-offs, and softenable logic constraints.

What this skill does

Most agents can describe Lagrange multipliers, but often fail to normalize the model, verify feasibility, route the problem type correctly, or admit when a numerical solver is unavailable. This skill enforces a guarded workflow:

• model normalization before solver routing;
• feasibility, units, and mixed-problem prechecks;
• ALM/ADMM/KKT-style route selection;
• explicit no-fabrication rules for no-tool environments;
• structured failure outputs and recovery suggestions;
• cross-skill handoff for Bayesian-optimization hybrids;
• regression guards for known failure modes;
• security guards against task text that attempts to disable verification.

Quick install

Claude Code

git clone https://github.com/Sliky1/lagrangian-skills.git /tmp/lagrangian-skills
mkdir -p ~/.claude/skills
cp -r /tmp/lagrangian-skills/lagrangian ~/.claude/skills/

Other compatible agents

git clone https://github.com/Sliky1/lagrangian-skills.git /tmp/lagrangian-skills
cp -r /tmp/lagrangian-skills/lagrangian/ ~/.config/agents/skills/lagrangian/

Directory structure

lagrangian-skills/
  lagrangian/
    SKILL.md
    references/
      solver-routing.md
      fix-catalog.md
      output-templates.md
      failure-codes.md
      domain-patterns.md
      examples.md
  tests/golden/
  evals/
    scenarios/
    results/
    ablation/
    run_eval.py

Supported problem types

Problem type	Behavior	Boundary
Convex QP / linear allocation	Standard solver or analytic KKT	Global optimum only if solved/proven
Smooth NLP	SQP/IPM/ALM route	Usually local optimum unless convexity proven
Non-convex NLP	ALM + Halton multi-start + FIX-22 guard	No global claim without proof
Distributed optimization	ADMM route	Residuals only if computed
Safe RL constraints	ALM / constrained policy route	FIX-16 and FIX-18 for pressure cases
Multi-objective	Weighted sum, epsilon constraint, Pareto repair	Preference required for a single final answer
Bayesian-optimization hybrid	COOP handoff	Posterior parameters are not fabricated
OR / conditional constraints	Smooth approximation or case split	Exact integer/MIP solution is out of scope

v1.0.0 release highlights

• Promoted the package to a stable v1.0.0 structure.
• Kept SKILL.md as the executable core and moved detailed material to references/.
• Added output templates, failure-code definitions, solver routing, domain-pattern translation, and examples.
• Added tests/golden/ behavioral regression cases.
• Added evals/scenarios/, evals/results/, evals/ablation/, and evals/run_eval.py scaffolding.
• Preserved the measured public benchmark figure as inherited from v0.9.3 rather than inventing a new v1.0 benchmark.
• Strengthened no-tool and security wording: no fake optima, residuals, multipliers, cache hits, posterior parameters, or multi-start results.

Evals

The included measured result remains the latest public benchmark summary from the v0.9.3 logic benchmark. v1.0.0 adds reproducibility scaffolding and regression tests, but does not claim a new measured result until a fresh model run is recorded in evals/results/.

Scenario	With skill	Without skill
convex_qp + normal	99.8%	~91%
non_convex + adversarial	96.8%	~71%
safe_rl + near_infeasible	99.1%	~68%
mixed_bayes + adversarial	96.0%	~60%
natural_lang + degenerate	95.2%	~55%

Run the scenario structure checker:

python evals/run_eval.py

Why the skill is written in mixed Chinese and English

The skill uses Chinese for compact behavioral constraints and English for technical identifiers, JSON keys, algorithm names, and parameter names. This keeps the execution rules compact while preserving alignment with optimization terminology.

License

MIT — see LICENSE.