forge-debate

name	forge:debate
description	Adversarial multi-model evaluation. Models argue for, against, and neutrally about a subject.
disable-model-invocation	true
argument-hint	[subject: path or proposal or instruction] [--output path] [--code] [--models m1,m2] [--worker model:stance]
context	fork
effort	high
allowed-tools	Bash, Read

Debate Evaluation

Run an adversarial multi-model evaluation where models argue for, against, and neutrally about a subject.

When invoked from Claude Code, execute the workflow now. Do not just restate these instructions, say "Command completed", or ask the user to run the commands manually unless a real prerequisite is missing.

Usage

/forge:debate [subject] [--code] [--models model1,model2]

Arguments

Argument	Required	Description
subject	Optional	File, directory, proposal, or instruction on what to evaluate (defaults to cwd)
--code	Optional	Switch: use code evaluation framework (default: proposal)
--models	Optional	Comma-separated model list (default: all available)
--worker	Optional	Repeatable: model:stance or model:"custom prompt" (mutually exclusive with --models)
--output	Optional	Write result to file instead of conversation (e.g., debate.md)

Available models: !forge workflow list-models

Only use models with status ready in the table above. If the default set includes unavailable models, pass --models <ready models> explicitly. If the user explicitly requested an unavailable model, stop and tell them what proxy or credential is missing rather than silently substituting. If no models are ready, tell the user what's missing and stop.

---

Execution

Step 1: Resolve Subject and Flags

Parse $ARGUMENTS into a positional subject and optional flags. The subject is everything that is not a recognized flag (file path, directory, proposal text, or free-form instruction). Strip any leading @ prefix on the subject. If no subject is found, default to the current working directory.

Recognized flags (extract from $ARGUMENTS if present):

• --code — switch
• --models <value> — comma-separated model list (mutually exclusive with --worker)
• --worker <value> — repeatable: model:stance or model:custom prompt
• --output <path> — write result to file instead of conversation

Never ask the user to clarify. If $ARGUMENTS contains anything, proceed immediately.

Step 2: Run Adversarial Evaluation

forge workflow debate "<subject>" [--code] [--models <models>] [--worker <spec>]... --json

Omit any flag the user didn't specify. Do not pass both --models and --worker.

Parse the JSON output. Each model receives a different stance (for/against/neutral) and evaluates the subject from that perspective. If the command fails, surface the real error and stop; do not claim success.

Step 3: Synthesize

Combine the perspectives:

1. Points of agreement across all stances
2. Key disagreements and which stance has stronger evidence
3. Risk assessment from the critic's perspective
4. Viability assessment from the supporter's perspective
5. Overall recommendation with confidence level

Make it clear which parts came from agreement across stances versus which parts remain disputed.

Output routing: If --output was specified, write the complete synthesis to that path using the Write tool (create parent directories if needed). Print a one-line confirmation: Wrote synthesis to {path}. Do not also print the full result in the conversation. If --output was not specified, print the result in the conversation as usual.

---

Models and Roles

Models are assigned stances cyclically. Default models (from forge workflow list-models):

Order	Default Model	Stance	Role
1st	gpt-5.5	FOR	Supporter -- strengths
2nd	gemini-2.5-pro	AGAINST	Critic -- risks
3rd	claude-opus	NEUTRAL	Analyst -- balanced view

Use --models to control which models participate. Stances cycle through for/against/neutral in order.

Code Mode

When --code is specified, models evaluate the target code from adversarial perspectives:

• FOR stance: Identifies good design, correct implementations, production readiness
• AGAINST stance: Identifies bugs, security issues, performance problems, architectural flaws
• NEUTRAL stance: Balanced assessment of code quality with file:line evidence

Requirements

• Forge CLI: forge must be on PATH
• Proxies: GPT-5.5 and Gemini require active proxies (forge proxy create litellm-openai)