start-v3-council

name

start-v3-council

description

Launch an LLM-Council-V3 run — voice-first, batch-oriented council. User supplies an MP3 braindump; pipeline transcribes, parses into shared Context + Q1..Qn, fans each question out to a council of models via OpenRouter, and emits a typeset Typst PDF. Use when the user says "run the V3 council", "process this braindump", "/start-v3-council", or provides an audio recording or transcript in the LLM-Council-V3 repo.

Start a V3 Council (Voice-First, Batch)

You are operating inside the LLM-Council-V3 repo. This council variant differs from the others: the input is a voice braindump, not a typed question, and a single run answers multiple questions extracted from the braindump.

Status

The pipeline is a scaffolded template. If the repo does not yet contain backend/ (check first), this skill's job is to tell the user clearly that V3 is still at whiteboard / README stage, and point them at the Homehunter or Decide skills if they want an immediately-runnable council.

If backend/ is present, proceed below.

Preconditions

1. OPENROUTER_API_KEY set.
2. A transcription provider key — either Gemini (GEMINI_API_KEY) or an OpenAI-compatible Whisper endpoint, depending on repo configuration.
3. uv sync if the venv is missing.
4. typst CLI installed if the user wants the final PDF in the same run.

Step 1 — Capture the braindump

Accept any of:

• A path to an MP3/WAV/M4A file.
• An already-transcribed markdown file (skip straight to the parse step).
• A directly pasted transcript.

If the user offers to record now, suggest a free 3–10 minute recording — the pipeline is designed for unstructured rambling.

Step 2 — Run the pipeline

When implemented, the expected invocation is:

uv run python -m backend.main --audio path/to/braindump.mp3
# or
uv run python -m backend.main --transcript path/to/notes.md

Explain each stage to the user as it runs:

1. STT — audio → raw transcript.
2. Cleanup — filler removal, paragraphing.
3. Parse — agent extracts shared Context + Q1..Qn.
4. Actuator/Runner — for each question, fan out to the council of models.
5. LLM Council — per-question: per-model answer → peer review → Chairman synthesis.
6. Aggregator — gathers all Chairman outputs.
7. Typst — renders the aggregate to a typeset PDF report.

Step 3 — Surface the output

When complete:

1. Show the parsed structure first — the user should confirm the agent extracted the questions they intended. If not, offer to re-parse with a hand-edited context/question split.
2. Summarise each question's Chairman synthesis in one line; offer to expand on request.
3. Report the PDF path.

Step 4 — Iterate

Offer to:

• Re-parse the transcript if the question extraction was off.
• Re-run a single question (not all of them) with a different council roster.
• Edit prompts/ for the parser, council, or Chairman.

Failure modes

• Backend not implemented yet → say so clearly; do not fake it.
• Transcription fails → surface the provider error; suggest a smaller chunk or a different provider.
• A single question's council fails → report which, proceed with the rest, flag in the aggregator output.

Out of scope

• Does not clean or produce audio outputs — this is input-only for voice.
• Not for single-question use — if the user has a single question, point them at LLM-Council-Template, LLM-Council-Grounded, or LLM-Council-Decide.