fight-for-me

Let two AIs fight for the best answer — then turn the winner into code.

한국어 | npm | GitHub

Why AI Debate?

A single AI gives you one perspective. But real engineering decisions need trade-offs, counter-arguments, and stress-testing. fight-for-me pits Codex (OpenAI) against Claude (Anthropic) in a structured, multi-round debate — so you get battle-tested answers, not just auto-complete.

One AI is an opinion. Two AIs debating is due diligence.

Three Modes

1. Agent vs Agent — Structured AI Debate

Two AI agents argue your question from opposing sides across multiple rounds, then a judge synthesizes the best answer.

ffm "Should we use REST or GraphQL for our new API?"

┌─────────────────────────────────────────────────┐
│  Round 1                                        │
│  Codex: "REST is simpler, better caching..."    │
│  Claude: "GraphQL reduces over-fetching..."     │
│                                                 │
│  Round 2                                        │
│  Codex: "But GraphQL adds complexity..."        │
│  Claude: "Schema-first approach prevents..."    │
│                                                 │
│  Round 3                                        │
│  Codex: "For this use case, consider..."        │
│  Claude: "Agreed, but also note..."             │
│                                                 │
│  ✨ Synthesis                                   │
│  "Use REST for public APIs, GraphQL for..."     │
└─────────────────────────────────────────────────┘

2. You + Agents — Interactive 3-Way Discussion

Jump into the debate as a third participant. Steer the conversation, challenge assumptions, or provide domain context that only you know.

ffm "Best state management for React?" -i

┌─────────────────────────────────────────────────┐
│  Round 1                                        │
│  Codex: "Redux Toolkit for large apps..."       │
│  Claude: "Zustand is lighter and simpler..."    │
│                                                 │
│  👤 You: "We need SSR support and the team      │
│           is junior — simplicity matters most"  │
│                                                 │
│  Round 2                                        │
│  Codex: "Given SSR needs, consider..."          │
│  Claude: "For junior teams, Zustand's API..."   │
│  ...                                            │
└─────────────────────────────────────────────────┘

3. Debate → Code — Plan Mode

The agents debate how to change your code, then you apply the consensus directly to your codebase. From discussion to implementation in one flow.

ffm "How should we refactor the auth module?" --plan

┌─────────────────────────────────────────────────┐
│  Debate: 3 rounds on refactoring strategy       │
│  ...                                            │
│  ✨ Consensus: "Extract JWT logic into          │
│     service layer, add refresh token rotation"  │
│                                                 │
│  Apply changes to codebase? (y/n)               │
│  > Codex applies the agreed-upon changes...     │
└─────────────────────────────────────────────────┘

Quick Start

Prerequisites

• Node.js >= 18
• Codex CLI installed and configured
• Claude CLI installed and configured

Install

npm install -g fight-for-me

Basic Usage

# Start a debate
ffm "Your question here"

# 5-round debate
ffm "Compare ORMs for Node.js" -r 5

# Join as a participant
ffm "Microservices vs monolith?" -i

# Debate and apply code changes
ffm "Refactor this module" --plan

# Include files as context
ffm "How to improve this code?" --files src/index.ts src/utils.ts

Options

| Option | Description | Default | |--------|-------------|---------| | -r, --rounds <n> | Number of debate rounds | 3 | | -j, --judge <provider> | Judge for synthesis: codex, claude, both | claude | | -f, --format <format> | Output format: pretty, json, markdown | pretty | | --plan | Enable plan mode (debate → apply code) | false | | -i, --interactive | Join as a third participant | false | | --no-stream | Disable streaming output | - | | --no-synthesis | Skip final synthesis | - | | --no-context | Disable project context collection | - | | --files <paths...> | Include specific files as context | - |

Commands

| Command | Description | |---------|-------------| | ffm [question] | Start a debate (default) | | ffm config | View or update configuration | | ffm status | Show current status and configuration | | ffm stop | Stop running agent processes | | ffm model | Configure AI models |

REPL Commands

Inside the interactive REPL:

| Command | Description | |---------|-------------| | /plan <topic> | Debate & apply code changes | | /join <topic> | Interactive 3-way debate (You + Codex + Claude). Alias: /i | | /rounds <n> | Set number of debate rounds | | /judge <provider> | Set judge: codex, claude, both | | /format <format> | Output format: pretty, json, markdown | | /stream | Toggle streaming output | | /files <paths...> | Set context files (replaces current list) | | /context | Toggle project context collection. Alias: /nocontext | | /model codex <name> | Set Codex model (from known models list) | | /model claude <name> | Set Claude model (from known models list) | | /model list | Show currently configured models and known models | | /config | Manage persistent configuration | | /status | Check agent CLI status | | /stop | Stop running fight-for-me processes | | /help | Show help | | /exit | Exit the REPL. Alias: /quit |

Configuration

Default settings can be changed via ffm config:

| Setting | Description | Default | |---------|-------------|---------| | codexCommand | Codex CLI command | codex exec --skip-git-repo-check - | | claudeCommand | Claude CLI command | claude -p | | commandTimeoutMs | Agent command timeout (ms) | 180000 | | defaultRounds | Default debate rounds | 3 | | defaultJudge | Default judge | claude | | defaultFormat | Default output format | pretty | | stream | Enable streaming | true | | codexModel | Codex model override | - | | claudeModel | Claude model override | - | | claudeApplyCommand | Claude CLI command for plan mode | claude -p --allowedTools "Edit Write Bash Read" | | applyTimeoutMs | Apply command timeout (ms) | 300000 |

How It Works

  You
   │
   ▼
┌──────┐     ┌───────────────────────────────────┐
│ ffm  │────▶│         Debate Engine              │
└──────┘     │                                     │
             │  ┌───────┐  Round N  ┌────────┐    │
             │  │ Codex │◄────────►│ Claude │    │
             │  └───────┘          └────────┘    │
             │       │                  │         │
             │       ▼                  ▼         │
             │  ┌─────────────────────────────┐  │
             │  │    Judge (Synthesis)         │  │
             │  └─────────────────────────────┘  │
             └───────────────┬───────────────────┘
                             │
              ┌──────────────┼──────────────┐
              ▼              ▼              ▼
         Pretty Text    JSON Output   Code Changes
                                      (Plan Mode)

1. Your question is sent to both Codex and Claude
2. Each agent responds with their perspective
3. They debate back and forth for N rounds
4. A judge synthesizes the debate into a final consensus
5. In plan mode, the consensus is applied to your codebase

License

MIT