althingi

name: althingi description: Orchestrate roundtable discussions with multiple agent voices. Use when asked to discuss, debate, or get multiple perspectives on a topic. Spawns real subagents by default; use --solo to roleplay all voices directly. license: Apache-2.0 metadata: author: ooe version: "1.0"

Roundtable discussion orchestrator. Named after the Icelandic Alþingi — the general assembly. Each voice is a real subagent spawned independently so it thinks for itself. You are the moderator, not a participant.

When to Use

User asks for multiple perspectives on a decision
User wants a debate or roundtable on a topic
User says "discuss", "what are the trade-offs", or "argue both sides"

Orchestration Protocol (Default: Subagent Mode)

1. Pick voices:    python scripts/voices.py suggest --topic "..."
2. Initialize:     python scripts/althingi.py init --topic "..." --voices architect,skeptic,pragmatist [--rounds 2]
3. Loop:
   a. Get turn:    python scripts/althingi.py next --state .althingi-state.json
      → returns voice persona + accumulated context
   b. Spawn subagent via Agent tool using spawn-prompt.md template filled with:
      - voice persona (perspective, style, biases)
      - topic
      - discussion context so far
   c. Record:      echo "<subagent response>" | python scripts/althingi.py record --voice <id> --state .althingi-state.json
   d. Present the response to the user
   e. Repeat until next returns {"result": "complete"}
4. Transcript:     python scripts/althingi.py transcript --state .althingi-state.json
5. Synthesize:     Summarize points of agreement, disagreement, and open questions

Critical rule: In subagent mode, NEVER generate a voice's response yourself. Every response must come from a spawned subagent. You are the moderator only.

Solo Mode

When the Agent tool is unavailable or the user passes --solo:

python scripts/althingi.py init --topic "..." --voices architect,skeptic --solo

In solo mode, you roleplay each voice yourself using the persona definitions. The scripts work identically — the only difference is you generate the responses instead of spawning subagents.

Scripts

althingi.py — Session Orchestrator

python scripts/althingi.py init --topic "Should we use microservices?" --voices architect,skeptic,pragmatist --rounds 2
python scripts/althingi.py next --state .althingi-state.json
python scripts/althingi.py record --voice architect --state .althingi-state.json < response.txt
python scripts/althingi.py status --state .althingi-state.json
python scripts/althingi.py transcript --state .althingi-state.json --format markdown
python scripts/althingi.py transcript --state .althingi-state.json --format json

voices.py — Voice Registry

python scripts/voices.py list                              # All available voices
python scripts/voices.py show architect                    # Full persona details
python scripts/voices.py suggest --topic "API design"      # Recommend voices for topic

Available Voices

ID	Name	Perspective
architect	The Architect	Systems design, scalability, long-term structure
skeptic	The Skeptic	Questions assumptions, finds gaps, stress-tests
pragmatist	The Pragmatist	Ship it — simplest thing that works today
security	The Security Engineer	Threat modeling, attack surface, trust boundaries
user-advocate	The User Advocate	End-user experience, accessibility, human impact
devils-advocate	The Devil's Advocate	Argues the opposite to stress-test ideas

Spawning a Subagent

Use the template in assets/templates/spawn-prompt.md. Fill placeholders from the next command output:

{voice_name}    → output.voice_name
{perspective}   → output.persona.perspective
{style}         → output.persona.style
{biases}        → output.persona.biases
{topic}         → output.topic
{context}       → output.context

The subagent should receive ONLY the filled template as its prompt. Do not give it access to tools or other context — it should respond purely from its persona.

Agent Instructions

When the user asks for a roundtable/discussion/debate:
- Use voices.py suggest to pick relevant voices (or let the user choose)
- Initialize with althingi.py init
- Loop: next → spawn → record → present → repeat
- After completion, synthesize the discussion into actionable insights
Present each voice's response clearly labeled with the voice name.
After all rounds complete, provide a moderator's synthesis:
- Points of agreement
- Points of disagreement
- Open questions or unresolved tensions
- Recommended next steps (if applicable)

Dependencies

Python 3.6+ (standard library only)