gods

Gods

Dynamic God Team discovery for Claude Code workflow and review skills.

This repository is a Claude Code plugin named gods. The primary command is:

/gods:review

The plugin does not hard-code a pantheon. It reads a skill or workflow prompt, extracts review axes, embeds those axes with candidate deity profiles, searches team sizes k=4..16, and writes the smallest powerful team that covers the workflow without wasteful overlap.

Commands

• /gods:review <task-or-skill-path> - discover and use a God Team for review.
• /gods:nihil <target-scope> - run the evaluated k=10 Nihil transformation workflow.
• /gods:forge - refine workflow axes or candidate god profiles.
• /gods:evolve - improve a previous run from its rejected gods and rejected sizes.
• /gods:rank <task-or-skill-path> - inspect scoring and rejected alternatives.

Local Usage

Install dependencies when needed:

python3 -m pip install -r requirements.txt

Run the default thermo-nuclear review example:

python3 main.py

Run against a specific skill markdown file or inline task:

python3 main.py --skill examples/thermo-nuclear-code-quality-review.md
python3 main.py --skill "Review a Python refactor for architecture, deletion, type boundaries, and approval bar."

Validate the plugin shape and generated artifacts:

python3 scripts/validate_gods_pipeline.py

Test the plugin in Claude Code:

claude --plugin-dir .

Then invoke:

/gods:review examples/thermo-nuclear-code-quality-review.md

Outputs

The discovery run writes:

results/god_team_scores.csv
results/god_team.json
visualizations/god_team_tsne.png

results/god_team.json contains the selected team, rejected gods, rejected team sizes, assigned workflow axes, and coverage/separation metrics.

The JSON includes the command, selected gods, role assignments, score breakdowns, evidence notes, and any blocked validations. The local pipeline uses deterministic TF-IDF vectors as a dry-run embedding backend, so no external API or Qdrant service is required for validation.

Model

The scoring model balances:

• semantic fit
• workflow coverage
• non-overlap
• mythic power
• name usability
• evidence score

t-SNE is only a visualization layer. It is not used as the final ranking metric.

See SEMANTIC_MODEL.md for the full model.