convez.ai
Skill
drgaciw

shiny-palm-tree

drgaciw/shiny-palm-tree

Ver no GitHub
InicialPluginScore · de 0 a 100

A suite of Claude Code **skills** (deterministic measurement scripts + guidance) and **subagents** (orchestration roles) that measure software-quality metrics on legacy Java codebases and turn the measurements into prioritized, pattern-mapped refactoring plans.

Instalação

$npx convez-ai drgaciw/shiny-palm-tree

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

Java Quality Metrics — Skills & Agents Suite

A suite of Claude Code skills (deterministic measurement scripts + guidance) and subagents (orchestration roles) that measure software-quality metrics on legacy Java codebases and turn the measurements into prioritized, pattern-mapped refactoring plans.

Core principle: scripts measure, the model interprets. Every metric is computed by a parser-based script (tree-sitter-java, no JVM); the LLM does prioritization and refactoring recommendation. All skills emit one common JSON schema so they compose without glue code.

Install

pip install -r requirements.txt
python tests/run_all.py      # 37 golden assertions + 2 integration tests

Layout

common/javametrics/      shared parser, kernels, schema, thresholds loader, CLI
common/thresholds.yaml    band boundaries + breach rules (one profile per codebase)
tools/merge_metrics.py    join multiple metric JSONs on (file, class, method)
skills/<name>/            SKILL.md + scripts/ + tests/ + reference.md
agents/<name>.md          subagent definitions
tests/run_all.py          full test battery

Skills

SkillGranularityOutput
cyclomatic-complexity-javamethodv(G) + size guards (loc, params, nesting)
halstead-metrics-javamethod/classVolume, Difficulty, Effort/Time/Bugs
cognitive-complexity-javamethodSonarSource cognitive complexity
ck-metrics-javaclassWMC, DIT, NOC, CBO, RFC, LCOM4 + class_loc
maintainability-index-javaclassMI (composes Halstead V + v(G) + LOC)
hotspot-analysisfilechurn × complexity rank, + JaCoCo coverage
refactoring-advisortargetsignature → pattern plan

Every measurement script runs standalone and emits the common schema:

python skills/cyclomatic-complexity-java/scripts/cyclomatic.py --path src/ --format json

Agents & orchestration flow

main session
  ├─▶ metrics-scanner    (read-only)  ─┐  run in parallel, shard large repos
  └─▶ hotspot-ranker     (read-only)  ─┘
            │
            ▼
      refactoring-planner  (plan mode, no writes)
            │
        human approval
            │
        implementation  (main session / ddd-refactoring)
            │
            ▼
      refactoring-verifier (read-only)  — re-measure, diff, confirm no regression

End-to-end pipeline (manual)

P=src/
python skills/cyclomatic-complexity-java/scripts/cyclomatic.py --path $P > vg.json
python skills/halstead-metrics-java/scripts/halstead.py        --path $P > hal.json
python skills/cognitive-complexity-java/scripts/cognitive.py   --path $P > cog.json
python skills/ck-metrics-java/scripts/ck.py                    --path $P > ck.json
python skills/maintainability-index-java/scripts/mi.py         --path $P > mi.json
python tools/merge_metrics.py vg.json hal.json cog.json ck.json mi.json > merged.json
python skills/hotspot-analysis/scripts/hotspot.py --repo . --metrics merged.json \
       --since "12 months ago" --coverage jacoco.csv > hotspot.json
python skills/refactoring-advisor/scripts/advise.py --metrics merged.json \
       --hotspot hotspot.json --format table

Design constraints

  • Metrics are a compass, not a gate. Skills report risk bands; they never auto-fail a build (Goodhart's law). The advisor flags when a metric is being gamed (e.g. method splitting that relocates complexity).
  • Thresholds live in thresholds.yaml, not in skill prose — calibrate per codebase. If >40% of methods land in the worst band, tune thresholds, don't mass-refactor.
  • Measurement agents are read-only. Only the human-approved refactoring loop touches source; the verifier re-measures afterward.
  • Known weak signals are demoted: Halstead Effort/Time/Bugs are indicative only; MI is a trend number (1990s regression constants), never an absolute grade.
  • Conventions are pinned in each skill's reference.md (Halstead operator/operand split, decision-point counting, cognitive increments) — golden fixtures encode hand-computed values so the conventions can't silently drift.
Ver o original no GitHub

Bom para

Relacionados