strg 0.0.1

Self-testing refactor generator — LLM agent pipeline that writes its own tests then optimises code

strg

LLM agent pipeline that writes its own tests, then optimises code against them.

Status: 0.0.1 — name-reservation placeholder. The real implementation lives in the llmios research repository today; the restructured src/strg/ package ships in 0.1.0.

What it will do (target: 0.1.0)

Four-phase pipeline driven by independent LLM agents that communicate through on-disk artefacts:

1. Profiling — rank the project's source files by runtime sensitivity.
2. Test generation — for each top-ranked target, an agent writes a pytest suite. Three sub-phases:
   • 2.1 Coverage: comprehensive baseline tests.
   • 2.2 Mutation: additional tests targeting surviving mutants from a fixed Cosmic Ray catalogue (Wong-Mathur selective operator set; 74 operators).
   • 2.4 Scaled: 1000×-scaled-input variants for benchmark sensitivity.
3. Combined-suite measurement — line/branch coverage, cyclomatic complexity, AST counts on the union of generated test files.
4. Improvement — a separate agent refactors the source for runtime while the test suite gates correctness.

Why "strg"

German keyboards label the Ctrl key "Strg" (Steuerung = control). The package controls / orchestrates LLM agents around your code. Three LLM-orchestration verbs fit the name as a backronym: Self-Testing Refactor Generator.

License

MIT — see LICENSE.