Ax GEPA For Python

Use when writing Python code with axllm for GEPA, Pareto tradeoffs, reflection clients, metric budgets, optimizer state, and artifacts.

Install

Install only this skill for Python:

Shell

npx skills add https://ax-llm.github.io/ax/python/ --skill 'ax-python-gepa'

Published skill file: ax-python-gepa/SKILL.md.

Source

Source: packages/python/skills/ax-python-gepa/SKILL.md
Version: 25.4.8

Skill Instructions

This skill helps an agent write Python code with the generated Ax package axllm. Use the generated package API, examples, and manifests; do not import TypeScript-only APIs unless you are editing the TypeScript package.

When To Use

Run the generated GEPA optimizer or inspect a GEPA artifact.
Use BootstrapFewShot before GEPA when demonstrations should seed optimization.
Track metric budgets, reflection calls, candidate state, and Pareto fronts.

Package Facts

Language: Python.
Package: axllm.
Package API docs: API.md and axir-api.json.
Capability manifest: axir-capabilities.json.
Runnable examples: examples/.
Real network support: yes.
Scripted no-key transport support: yes.
Runtime profiles: javascript-quickjs, python-pyodide.

Core Pattern

Python

from axllm import AxGEPA

engine = AxGEPA(reflection_client)
result = engine.optimize(request, evaluator)

Relevant API Surface

Optimizers: optimize, playbook, AxPlaybook, AxBootstrapFewShot, AxGEPA, OptimizerEngine, OptimizerEvaluator

Guardrails

Start from package examples for exact native syntax before inventing a new call shape.
Use provider-api examples only when the user explicitly has provider credentials available.
Use no-key examples for deterministic local checks and provider request mapping.
Treat AxIR as the source of generated package truth: if package docs disagree with source code, update the compiler and regenerate packages.
Do not copy repo-maintainer skills from tools/*/skills/ into user packages.
GEPA returns baselineScore and baselineScores at the artifact top level and in metadata, measured before bootstrap or component changes. Compare metadata.bestScore with baselineScore; the original candidate need not remain on the Pareto frontier.
Baseline and seeded validation use the same capped held-out subset and selection scalarization. Both consume maxMetricCalls. Bootstrap generation has separate limits; metadata.bootstrapMetricCalls plus metadata.totalMetricCalls equals metadata.totalCalls. Without new demos, reuse the baseline evaluation.
Bootstrap demos contain only fresh training inputs and outputs; never recover demos from a retained validation or inactive-child trace tree.