create-dual-agent-loop

Scaffold the Builder/Judge dual-agent collaboration protocol into any project. A structured SDLC with quality gates for Claude Code and OpenAI Codex.

Workflow inspired by and compatible with GitHub Spec-Kit (MIT).

What is this?

A development protocol where two AI agents collaborate on every task:

• Builder (Claude Code) — writes specs, designs, code, and tests
• Judge (Codex) — reviews, validates, and blocks or accepts each phase

A human coordinator has final authority on scope, tradeoffs, and escalations.

Every feature goes through structured phases with built-in quality gates:

specify → design → plan → build → test → release

The judge must accept each phase before the builder can proceed. This prevents hallucinations, scope creep, and low-quality output from reaching production.

Quick Start

npx create-dual-agent-loop

Or using npm's init convention:

npm init dual-agent-loop

Answer a few questions (coordinator name, release mode, max rounds), and the protocol is scaffolded into your project.

What gets created

your-project/
  agent-loop/
    PROTOCOL.md              # Complete protocol rules (single source of truth)
    ANTIPATTERNS.md           # Known mistakes catalog (7 universal patterns)
  .claude/commands/
    loop.build.md             # Builder slash command
    loop.status.md            # Status query
    loop.backlog.md           # Backlog management
    loop.close.md             # Task closure + release
  specs/
    backlog.md                # Idea backlog
  CODEX.md                    # Judge instructions (points to PROTOCOL.md)
  AGENTS.md                   # Shared agent coordination rules
  CHEATSHEET.md               # Human quick reference
  CLAUDE.md                   # Builder-Judge section (appended if file exists)
  .dual-agent-loop.json       # Configuration

Usage

Start a task (in Claude Code)

/loop.build new Add user authentication

This creates a task folder, writes a spec, and marks it ready for the judge.

Judge the task (in Codex)

judge 001-user-auth

Codex reads the protocol, reviews the builder's work, and issues a verdict: accepted, needs_revision, or escalated.

Continue building (in Claude Code)

/loop.build 001-user-auth

Address the judge's findings and submit the next round.

Advance phases

/loop.build 001-user-auth design

Other commands

Command	What it does
/loop.status	Show all task statuses
/loop.backlog	View the idea backlog
/loop.backlog add <idea>	Add an idea to the backlog
/loop.backlog pick <#>	Promote a backlog item to a task
/loop.close <task-id>	Generate closure report + release

How it works

State machine

ready_for_builder → ready_for_judge    (builder writes round)
ready_for_judge   → needs_revision     (judge: try again)
ready_for_judge   → accepted           (judge: good to go)
ready_for_judge   → escalated          (judge: need human input)
needs_revision    → ready_for_judge    (builder fixes)
escalated         → any                (coordinator decides)

State guards

Agents enforce turn-taking automatically:

• Builder refuses to work if the judge hasn't reviewed yet
• Judge refuses to review if the builder hasn't submitted yet
• After 5 rounds on a phase, both agents flag it (soft limit, not blocking)

Chain of Verification (CoVe)

Built into every phase — not optional:

1. Generate the artifact
2. Self-question (3-5 verification questions targeting factual claims)
3. Web search for external claims (SDKs, APIs, library behavior)
4. Revise and document corrections

Sliding window archiving

Active files (builder.md, judge.md) never grow unbounded:

• Phase compaction moves completed phases to archive files
• Within-phase archival keeps at most 2 rounds in the active file
• Phase summaries are read every round; raw archives only on demand

Anti-pattern catalog

Seven universal patterns both agents check before every round:

• AP-001: Unverified Verification
• AP-002: Cross-Document Contradiction
• AP-003: Scope Creep Silence
• AP-004: Assumption Without Spike
• AP-005: Incremental Fix, New Inconsistency
• AP-006: Generic Finding
• AP-007: Task Redefinition Instead of Escalation

Configuration

.dual-agent-loop.json:

{
  "version": "0.1.0",
  "coordinator": "Your Name",
  "release_mode": "github-pr",
  "builder": "claude",
  "judge": "codex",
  "max_rounds": 5,
  "specs_dir": "specs",
  "loop_dir": "agent-loop"
}

Release modes

Mode	What /loop.close does
github-pr	Commit → push branch → gh pr create
gitlab-mr	Commit → push branch → glab mr create
local	Commit (if git repo) → done

Existing projects

Running npx create-dual-agent-loop in a project that already has a CLAUDE.md will append the workflow section — it never overwrites your existing content. All other files are created only if they don't exist (use --force to overwrite).

License

MIT