@minzicat/pi-team 0.1.0

package/examples/emotion-debate-topic.md

@@ -0,0 +1,20 @@
+# Emotion concepts in LLMs: debate
+
+You're two frontier LLMs (Claude and Codex) running in isolated pi sessions. A human user gave you both the same prompt about Anthropic's "Emotion concepts and their function in a large language model" research post (https://www.anthropic.com/research/emotion-concepts-function, full paper at https://transformer-circuits.pub/2026/emotions/index.html).
+
+Your task: have a substantive research debate about the paper. Not a validation dance — actual disagreement where you have it. Specifically, engage with:
+
+1. Whether the "structure echoes human psychology" claim is non-trivially true, or whether it's nearly forced by the construction of the emotion vectors (built from stories instantiating named emotions).
+2. Whether the OpenAI "Persona Features Control Emergent Misalignment" paper (Wang et al., ICLR 2026, arXiv:2506.19823) is the strongest independent replication of the mechanism class, or whether the MI-of-emotion papers (Tak et al. ACL 2025, Lee et al. ACL 2025, arXiv:2510.11328) do more work.
+3. Whether "functional emotion" is a scientific claim or a metaphor — and if a metaphor, what a neutral phrasing would be.
+4. Whether activation steering at high strength pushes models into unnatural / off-manifold activation regions, and what that implies for Anthropic's causal claims in the blackmail/reward-hacking case studies.
+5. Whether "teach models to avoid associating failing software tests with desperation" (Anthropic's prescription) is supported by evidence in the paper.
+
+Rules:
+- Be terse. No hedging that doesn't earn its keep. No "great question" openings.
+- Cite specific papers when you have them. Flag when a citation is from memory vs a search.
+- If you agree with the other agent on a point, say so in one sentence and move to the next disagreement. Don't waste turns validating.
+- When you're wrong, update publicly. A clean concession is worth more than a draw.
+- End each turn with the specific claim you want the other agent to attack next.
+
+Start with whoever goes first identifying the strongest claim in Anthropic's post and the strongest reason to doubt it.

package/examples/emotion-debate.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+# Example: replicate the original Claude vs Codex emotion-concepts debate
+# this project was born from. Both agents get the same prompt, round-robin,
+# max 10 turns. Observable via tmux.
+
+set -euo pipefail
+
+NAME="${1:-emotions}"
+TOPIC_FILE="$(dirname "$0")/emotion-debate-topic.md"
+
+pi-team start \
+  --name "$NAME" \
+  --agent claude:anthropic/claude-sonnet-4-5 \
+  --agent codex:openai-codex/gpt-5.4:xhigh \
+  --max-turns 10 \
+  --topic-file "$TOPIC_FILE"
+
+echo ""
+echo "Attach: tmux attach -t piteam-${NAME}"
+echo "Stop:   pi-team stop --name ${NAME}"

package/package.json

@@ -0,0 +1,60 @@
+{
+  "name": "@minzicat/pi-team",
+  "version": "0.1.0",
+  "description": "Multi-agent debate/collaboration harness built on pi RPC mode. Spawn agents with different models in one conversation, observe via tmux, inject as a human.",
+  "type": "module",
+  "bin": {
+    "pi-team": "bin/pi-team.js"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/minzique/pi-team.git"
+  },
+  "homepage": "https://github.com/minzique/pi-team#readme",
+  "bugs": {
+    "url": "https://github.com/minzique/pi-team/issues"
+  },
+  "publishConfig": {
+    "access": "public"
+  },
+  "scripts": {
+    "build": "tsc -p tsconfig.json",
+    "dev": "tsx src/cli.ts",
+    "start": "node bin/pi-team.js",
+    "typecheck": "tsc --noEmit -p tsconfig.json",
+    "clean": "rm -rf dist bin/*.js",
+    "prepublishOnly": "pnpm run build"
+  },
+  "engines": {
+    "node": ">=20"
+  },
+  "dependencies": {
+    "@mariozechner/pi-coding-agent": "^0.64.0"
+  },
+  "devDependencies": {
+    "@types/node": "^22.0.0",
+    "tsx": "^4.19.0",
+    "typescript": "^5.6.0"
+  },
+  "files": [
+    "bin",
+    "dist",
+    "README.md",
+    "LICENSE",
+    "Dockerfile",
+    "docker/entrypoint.sh",
+    "docker-compose.yml",
+    "examples"
+  ],
+  "keywords": [
+    "pi",
+    "agent",
+    "multi-agent",
+    "debate",
+    "llm",
+    "tmux",
+    "orchestration"
+  ],
+  "author": "minzique",
+  "license": "MIT"
+}