dynamic-skill-compiler 0.1.0

Task-driven compiler for agent skill libraries.

Dynamic Skill Compiler

Dynamic Skill Compiler (DSC) is a task-driven compiler for agent skills. It takes a natural-language task, retrieves candidate skills from a local skill library, builds a skill graph, and emits a compact, dependency-aware skill package for execution.

What DSC Does

• Decomposes a task into subgoals and required capabilities.
• Retrieves local skill assets and their declared relationships.
• Extracts relevant skill fragments instead of passing whole skill files.
• Builds a graph over similar_to, belong_to, compose_with, and depend_on relationships.
• Runs compiler passes to select, repair, augment, and prune the skill set.
• Produces a CompiledSkillPackage with selected skills, execution order, coverage metrics, and pass traces.

Repository Layout

src/dynamic_skill_compiler/   Core DSC compiler package
tests/                        Compiler and experiment integration tests
experiments/                  Benchmark runners and local skill libraries
reports/                      DSC design notes

The benchmark runners under experiments/ are kept as evaluation harnesses for ScienceWorld, ALFWorld, and WebShop. They are intentionally separate from the core compiler package.

Quick Start

Install from a local checkout:

python -m pip install -e .

After the package is published to PyPI, users can install it with:

python -m pip install dynamic-skill-compiler

Semantic embedding support is optional:

python -m pip install "dynamic-skill-compiler[semantic]"

from dynamic_skill_compiler import (
    CompilerConfig,
    DynamicSkillCompiler,
    LocalEnvironment,
    LocalSkillLibraryRetriever,
)

retriever = LocalSkillLibraryRetriever("experiments/src/skills/scienceworld")
compiler = DynamicSkillCompiler(
    retriever=retriever,
    config=CompilerConfig(),
)

compiled = compiler.compile(
    "Your task is to measure the temperature of an unknown substance.",
    environment=LocalEnvironment(benchmark="scienceworld"),
)

print([skill.asset.name for skill in compiled.compiled_skills])
print(compiled.execution_order)
print(compiled.metrics.coverage_score)

CLI

The package installs a dsc command for quick local compilation:

dsc "Measure the temperature of the unknown object." \
  --skills-dir experiments/src/skills/scienceworld \
  --pretty

The command prints a JSON summary containing selected skills, execution order, coverage metrics, compiler pass traces, and dropped-skill reasons.

Build And Publish

Build source and wheel distributions:

python -m pip install ".[dev]"
python -m build
python -m twine check dist/*

Publish to TestPyPI first:

python -m twine upload --repository testpypi dist/*

Publish to PyPI:

python -m twine upload dist/*

Use a PyPI API token via TWINE_USERNAME=__token__ and TWINE_PASSWORD=pypi-... or through your local .pypirc.

Tests

PYTHONPATH=src python -m unittest discover -s tests -p "test_*.py"

PYTHONPATH=src:experiments/src .venv-experiments/bin/python \
  experiments/test_runtime_execution_guard.py

Branches

• main: standalone DSC codebase.
• codex/v0321: latest DSC development branch mirrored into main.