forge

Blog Post · Architecture Deep Dive · Tracing

Adversarial multi-agent pipeline for Claude Code. Separate AI agents generate and critique each other's work in adversarial feedback loops, where generators produce artifacts, discriminators validate them, and iteration drives convergence. Each agent runs in its own context window with fresh perspective.

Install

Plugin (marketplace):

/plugin marketplace add hatmanstack/claude-forge
/plugin install forge@claude-forge
/reload-plugins

The first command registers the marketplace (persisted to ~/.claude/plugins/known_marketplaces.json, so you only do it once). The second opens the install TUI — select a scope and confirm. The third activates the plugin in your current session.

When installed as a plugin, skills are prefixed with forge: — e.g. /forge:pipeline, /forge:brainstorm. The unprefixed forms (/pipeline, etc.) shown in the usage examples below apply to the standalone install path.

Standalone (copy into any project):

cp -r skills/ /path/to/your-project/.claude/skills/
cp -r agents/ /path/to/your-project/.claude/agents/
# Or personal (all projects)
cp -r skills/ ~/.claude/skills/
cp -r agents/ ~/.claude/agents/

Copy both skills/ and agents/ — the pipeline roles are native Claude Code subagents that live in agents/. When installed standalone, the orchestrator addresses them without the forge: plugin prefix (e.g. planner instead of forge:planner).

Requires Claude Code v1.0.33+ and a git-initialized project.

Enable Agent Teams (Required)

Claude Forge relies on the Agent and SendMessage tools for multi-agent orchestration. These now require an experimental feature flag.

Set the environment variable before launching Claude Code:

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

To make it permanent, add that line to your shell profile (~/.bashrc, ~/.zshrc, etc.) and restart your terminal.

Without this flag, skills that spawn or communicate with sub-agents will fail.

Skills

Skill	Purpose	Output	Next Step
brainstorm	Interactive design session, explores codebase, asks scoping questions	brainstorm.md	pipeline
audit	Combined audit runner, select any combination of eval, health, docs	Multiple intake docs	pipeline
repo-eval	3-evaluator panel scoring 12 pillars	eval.md	pipeline
repo-health	Technical debt audit across 4 vectors	health-audit.md	pipeline
doc-health	Documentation drift detection across 6 phases	doc-audit.md	pipeline
pipeline	Automated build/remediation cycle, routes by intake doc type	Committed code	Done

Usage

# Feature development
/brainstorm I want to add webhook support for payment events
/pipeline 2026-03-12-payment-webhooks

# Full audit (health > eval > docs) with one pipeline run
/audit all
/pipeline 2026-03-15-audit-my-app

# Or run individual audits (each creates its own plan directory)
/repo-eval
/pipeline 2026-03-15-eval-my-app

Resume any interrupted pipeline by re-running /pipeline with the same slug.

Pipeline Flows

Feature (brainstorm.md)

Planner ↔ Plan Reviewer → Implementer ↔ Code Reviewer → Final Reviewer
         (max 3 iter)                   (max 3 iter/phase)    GO/NO-GO

Repo Eval (eval.md)

3 Evaluators → Planner ↔ Plan Reviewer → Implementer ↔ Reviewer → Verify
(parallel)     (max 3)                   (max 3/phase)             verify findings

Repo Health (health-audit.md)

Auditor → Planner ↔ Plan Reviewer → Hygienist ↔ Health Reviewer → Fortifier ↔ Health Reviewer → Verify
                                     [cleanup]                      [guardrails]                   verify findings

Doc Health (doc-audit.md)

Doc Auditor → Planner ↔ Plan Reviewer → Doc Engineer ↔ Doc Reviewer → Verify
                                         [fix + prevent]               verify findings

File Structure