Vega Labs

Vega Labs is the core-x observatory for repositories, libraries, models, skills, and reusable capabilities. It turns GitHub stars and owned repositories into searchable intelligence, canonical skills/rules/flows, adoption decisions, technology alerts, and draft-only maintenance actions.

The app is local-first. The runtime contract is OpenResponses, and local mode defaults to /bus, proxied to the MLX Gateway at http://127.0.0.1:8090. Chat and reasoning use /v1/responses; screenshots, PDFs, diagrams, and tables use OCR at /v1/ocr/*. Web mode exists for deployed testing only.

Operating Model

• News is a hybrid intelligence surface for watched repo updates, adoption radar, research queue state, skill extraction opportunities, and next actions.
• Mine is the owned-repo execution workspace for README gaps, maintenance, deployment/test signals, package plans, template extraction, and model-specific mission briefs.
• Ops is the draft-only inbox for action items generated by daily ops, weekly research, or manual review.
• ReadmePanel and ChatPanel route through typed Vega Labs tools instead of one-off prompts.
• Codex, Claude, and local MLX mission briefs are adapter outputs from the same canonical extraction record.
• Ops kits generate draft README, AGENTS, maintenance, deployment, testing, and action-item artifacts from evidence.

Canonical Artifacts

Raw snapshots:

• data/data.json
• data/my-repos.json
• mirrored in public/ for the UI

Derived intelligence:

• repo-signals.json
• research-queue.json
• skill-extractions.json
• mine-health.json
• repo-inspections.json
• action-items.json
• automation-runs.json
• ops-digest.json
• weekly-research-review.json
• template-kits.json
• repo-ops-kits.json

Every derived file is mirrored in both data/ and public/.

Tracked template sources live under templates/ and cover house references, repo Ops kits, core-x ecosystem context, good-candidate criteria, and the local-first MLX/OpenResponses policy.

Tools

Vega Lab exposes MCP/OpenResponses tools from src/mcp-server/index.js, including discovery, statistics, research queue, adoption ranking, skill extraction, template kit listing, Ops kit generation, mission generation, repo inspections, action inbox updates, ops digest generation, weekly research review, runtime health, model-zoo inspection, and OCR-backed visual/PDF evidence extraction.

The MCP server identity is vega-lab-mcp. Legacy git-stars:* keys are read for one migration cycle where local compatibility is required.

Review Gate

Vega Labs outputs are reviewable drafts by default. Raw repo notes can remain internal, generated SKILL candidates default to internal/pending, public summaries require review, and no workflow may automatically publish, open PRs, merge, deploy, create new houses, or write outside approved Vega output folders.

Local Development

pnpm install
pnpm run dev

Optional data refresh:

GITHUB_TOKEN=<your_pat> pnpm run build:data
GITHUB_TOKEN=<your_pat> pnpm run sync:mine
pnpm run generate:stats
pnpm run test:data

The nightly GitHub Action requires a repository secret named GH_PAT. Use a personal GitHub token with repo access so Vega Lab can refresh public stars plus owned, private, collaborative, and organization repositories. Do not rely on the default github.token for Mine data; it authenticates as GitHub Actions and cannot represent the KBLLR account.

Verification:

pnpm test
pnpm run build

Local Runtime

Start the MLX/OpenResponses Gateway separately, then run the app. Vite proxies /bus to:

http://127.0.0.1:8090

Override the proxy target with VITE_EVENT_BUS_PROXY_TARGET or VITE_LOCAL_OPENRESPONSES_URL when the gateway runs elsewhere.

Runtime Settings shows the active mode, model, bus URL, tool availability, and failure state. Local mode is the default expected path.

OCR evidence routes through the gateway at /bus/v1/ocr/* and delegates to the local mlx-vision worker. OCR outputs stay reviewable drafts and should be attached to repo analysis, skill evidence, action items, or Ops kits only after review.

Automations

• Daily Ops refreshes snapshots, inspects owned repositories, generates ops-digest.json, and drafts action inbox items.
• Weekly Research ranks bright-star candidates, identifies reusable skills/rules/flows, generates weekly-research-review.json, and drafts recommendations.

Automation authority is draft-only. Vega Lab may generate inbox items, Ops kits, and mission briefs, but it must not open PRs, merge, deploy, or mutate other repositories without explicit human action.