forge-council

forge-council

A single AI agent is a single perspective. It gives you one take — its own — and misses everything outside its frame. forge-council provides specialist agents that work from independent perspectives: architecture, design, implementation, testing, security, product, research, hiring. Assemble them into councils for multi-round debates, or invoke any specialist standalone.

Quick Start

git clone https://github.com/N4M3Z/forge-council.git
cd forge-council
make install     # forge install — deploys agents, skills, rules to all providers

Then in your session:

/DebateCouncil [topic]
/DeveloperCouncil [task]
/ProductCouncil [requirements]
/KnowledgeCouncil [knowledge-management topic]
/HiringCouncil [job posting or role design]

Or invoke any specialist standalone — no council needed:

Task: TheOpponent — "We should rewrite the backend in Rust"
Task: WebResearcher — "Best practices for rate limiting in distributed systems"
Task: SecurityArchitect — "Threat model our authentication system"

Note: make install deploys to the local workspace (./.claude, ./.gemini, ./.codex, ./.opencode). To install at user scope, run forge install --target ~.

Makefile Commands

make install     # forge install — assemble + deploy for every provider
make validate    # pre-commit cascade (prek → forge validate → validate.sh fallback)
make clean       # remove build/ artifacts

What it does

3-round debate — All councils use a structured debate where specialists respond to each other's points across three rounds: initial positions, challenges, convergence. The lead synthesizes areas of agreement, remaining disagreements, and recommended actions.

Council skills — /DeveloperCouncil for code review, architecture, and debugging. /DebateCouncil for cross-domain strategy and design debates. /ProductCouncil for requirements, features, and go/no-go decisions. /KnowledgeCouncil for knowledge architecture and memory lifecycle decisions. Each selects the right specialists for the task.

User checkpoints — After Round 1, the lead shows you the initial positions and asks for your input before the debate rounds begin. Add context, redirect focus, or skip to synthesis. Override with autonomous (no stops), interactive (stop every round), or quick (one round only).

Standalone specialists — Every agent works independently via the Task tool. TheOpponent stress-tests your proposals. WebResearcher investigates topics across the web. No orchestration needed.

What it looks like

You have an idea for a product. Before you write a line of code, you put it in front of the council:

/DebateCouncil We're building an AI code review tool for enterprise — $50/seat,
targeting engineering teams of 50+. Should we go all-in?

Round 1 — Initial Positions:
  SystemArchitect: The hard part isn't the AI — it's the integration surface.
              You need to plug into GitHub, GitLab, Bitbucket, and on-prem.
              That's four integration contracts before you ship one review.
  UxDesigner: Enterprise buyers won't even demo without SSO. Your landing
              page needs a "Book a demo" button, not a "Sign up" button.
              The first user isn't a developer — it's a procurement team.
  SoftwareDeveloper: $50/seat is a race to the bottom — GitHub Copilot is $19.
              Sell saved engineering hours, not seats. "This tool caught
              42 bugs last sprint" beats a per-seat invoice.
  WebResearcher: Three competitors pivoted from per-seat to usage-based pricing
              in 2025. CodeRabbit, Sourcery, and Codium all moved to
              per-PR or per-repo models after enterprise pushback on seat
              math for a tool that reviews code, not writes it.

[Checkpoint: Any context to add before debate?]

Round 2 — Challenges:
  SystemArchitect: "WebResearcher's data on pricing pivots changes the equation.
              If we go usage-based, metering infrastructure becomes day-1
              architecture, not a nice-to-have."
  UxDesigner: "SoftwareDeveloper's right on the value pitch, but procurement
              still needs a number. Usage-based with a monthly cap gives
              both — predictable budgets, honest pricing."
  SoftwareDeveloper: "SystemArchitect's integration concern is the real blocker.
              Start with GitHub only. One integration done well beats four
              done poorly. Validate with 10 teams before touching GitLab."
  WebResearcher: "SoftwareDeveloper's 'GitHub first' aligns with the data — 73%
              of enterprise teams use GitHub. But SystemArchitect should flag: the
              GitHub API rate limit (5000/hr) will bottleneck teams over
              200 developers."