start-grounded-council

name

start-grounded-council

description

Launch an LLM-Council-Grounded run — six personality-based council members deliberate with retrieved context (Pinecone + Tavily). Use when the user says "run the grounded council", "deliberate on this with research backing", "/start-grounded-council", or provides a question in the LLM-Council-Grounded repo. Handles backend + frontend startup, confirms grounding configuration, and surfaces the chairman synthesis.

Start a Grounded Council

You are operating inside the LLM-Council-Grounded repo. This skill runs a deliberation with retrieved context from Pinecone and/or Tavily.

Preconditions

1. Repo root must contain backend/main.py, main.py, and start.sh.
2. Check env: OPENROUTER_API_KEY is required. Grounding layers are optional but have their own keys:
   • TAVILY_API_KEY — for web/news grounding.
   • PINECONE_API_KEY + index config — for knowledge-base grounding. If a grounding key is missing, tell the user what that disables; do not invent values.
3. uv sync and cd frontend && npm install && cd .. if dependencies are not installed.

Step 1 — Capture the question

Ask the user for the question they want the council to deliberate on. One question per run.

Help the user judge whether grounding is actually needed:

• Pure conceptual / philosophical → no grounding needed; the planning agent will confirm.
• Factual with time-sensitive elements → Tavily useful.
• Domain-specific (against a Pinecone KB the user has configured) → Pinecone useful.

Step 2 — Start the app

Preferred:

./start.sh

Or manually (two terminals):

# Terminal 1
uv run python -m backend.main
# Terminal 2
cd frontend && npm run dev

Then direct the user to http://localhost:5173.

If the user wants a headless CLI run without the frontend, check whether main.py supports that directly — if yes, use it; if not, tell the user the current UX is the web UI.

Step 3 — Surface the output

The web UI shows:

• Planning agent's grounding choice (which sources were queried and why).
• Six per-personality responses with retrieved context visible.
• Peer-review rankings.
• Chairman synthesis.

If the user wants the PDF / podcast digest, trigger those from the UI, then direct them to the output paths. The typst-council-report skill can render a PDF from a captured council result JSON.

Step 4 — Iterate

Offer to:

• Re-ask with different grounding scope (e.g., "force Pinecone only").
• Edit personality prompts in backend/ if a member is drifting off their assigned style.

Failure modes

• Missing OPENROUTER_API_KEY → stop.
• Frontend port collision → offer to run on an alternate port.
• Pinecone index not found → confirm the index name in config; do not auto-create indexes.

Out of scope

This skill does not ingest or manage the Pinecone knowledge base — that is a separate operation. Use Daniel's pinecone MCP tools or the repo's ingestion scripts.