Vybestack Site

Static marketing site for vybestack.dev plus the vybescore/ dashboard that visualizes LLxprt evaluation results.

Structure

• index.html, vybestack.css, assets/: main landing page
• blog/: markdown posts + rendered HTML (kept up to date via .github/workflows/build-blog.yml)
• vybescore/: published VybeScore dashboard generated from the acoliver/evals repo

Publishing the VybeScore Dashboard

The workflow in .github/workflows/publish-vybescore.yml runs every 6 hours (and on manual dispatch) to:

1. Check out this repo and acoliver/evals
2. Install the latest @vybestack/llxprt-code CLI plus evaluation dependencies
3. Run npm run eval:all inside the evals repo using env-driven LLxprt configs
4. Build the dashboard artifacts via npm run build:vybes
5. Merge the previously published dashboard (copied from vybescore/) with the new results, ensuring historical runs stay in vybes-*.json, and sync the refreshed public/ folder to vybescore/ before committing

Required Secrets

Configure these repository secrets before enabling the workflow:

Secret	Purpose
SYNTHETIC_KEY	API key for the Synthetic GLM 4.6 provider (fed to llxprt --key)
CEREBRAS_KEY	API key for the Cerebras Qwen 3 provider
ZAI_KEY	API key for the ZAI GLM 4.6 provider

Optional Variables

Custom Repository variables can override non-sensitive CLI args, e.g. SYNTHETIC_BASEURL, CEREBRAS_MODEL, etc. Defaults mirror the values described in acoliver/evals/README.md.

Manual publish

Kick off Publish VybeScore from the Actions tab (workflow_dispatch) after updating evals or secrets. The dispatch form lets you choose a single LLxprt config if you need to re-run a specific provider.

Local publish helper

Use scripts/publish-vybescore.sh to run the same sequence locally:

export SYNTHETIC_KEY=...
export CEREBRAS_KEY=...
export ZAI_KEY=...

# assumes acoliver/evals is checked out at ../doevals/evals
./scripts/publish-vybescore.sh

Environment variables:

Variable	Description
EVALS_DIR	Path to the acoliver/evals checkout (defaults to ../doevals/evals)
LLXPRT_CONFIGS	Comma-separated config list (defaults to all LLxprt profiles)
LLXPRT_EVAL_ARGS	Extra flags appended to npm run eval:all (e.g. --max-passes 1)
SKIP_EVALS	Set to 1 to skip running npm run eval:all and only rebuild/publish

After the script finishes, review and commit the refreshed vybescore/ artifacts.