neural-codex

Codex-native prompts, templates, scripts, and agents that bring the neural-claude workflow to the Codex CLI. Everything is file-based, repo-local, and designed for repeatable iteration with clear state.

• No Claude-specific hooks, status lines, or TTS
• All state lives in plans/ and .codex/
• Prompts are namespaced as neural.*

What you get

Prompts

• Loop control: neural.loop-start, neural.loop-plan, neural.loop-status, neural.loop-cancel
• Planning: neural.plan, neural.plan-execute
• Memory: neural.memory, neural.recall
• Routing & analysis: neural.route, neural.question, neural.pv, neural.evolve
• Research: neural.research, neural.gh-learn, neural.yt-learn
• Sync & changelog: neural.sync, neural.changelog-architect
• Task tracking: neural.todo-new, neural.todo-check
• Meta creation: neural.meta.agent, neural.meta.skill, neural.meta.prompt, neural.meta.improve, neural.meta.eval, neural.meta.brain
• Output styles: neural.output-style (default/concise/table/yaml/html/genui)
• Skills & config: neural.skill, neural.profile, neural.test

Skills

Project-scoped skills in .codex/skills/:

• autonomous-loop: Ralph loop usage and guardrails
• worktree-manager: parallel worktrees for multi-session work
• code-reviewer: production-minded reviews
• memory-system: progress-log memory
• pattern-detector: PRD/progress pattern analysis
• prompt-engineering: prompt creation/refinement
• plan-execute: structured planning and execution
• youtube-learner: transcript-based summaries
• skill-creator: bootstrap new skills with SKILL.md template
• skill-installer: install external skills from URLs/registries
• deep-research: multi-source comprehensive research
• test-runner: smart test execution with Ralph integration

Templates

• plans/prd.json and plans/progress.jsonl
• expertise.template.yaml
• todo-workflow.md

Scripts

• scripts/ralph-loop.sh and scripts/ralph-once.sh
• scripts/memory_read.py / scripts/memory_write.py
• scripts/youtube-transcript.py
• scripts/setup-global.sh / scripts/setup-project.sh

Agents

• agents/multi-ai/AGENTS.md
• agents/dispatcher/AGENTS.md
• agents/meta-agent/AGENTS.md

Quick setup

1) Run the global install from this repo:

scripts/setup-global.sh

2) Restart Codex so /prompts:neural.* are picked up. 3) In any project, run the project install:

scripts/setup-project.sh

4) Verify prompts:

/prompts:neural.loop-start

Loop prerequisites

The Ralph loop requires flock and timeout.

macOS (Homebrew):

brew install util-linux coreutils
export PATH="/opt/homebrew/opt/util-linux/bin:/opt/homebrew/opt/coreutils/libexec/gnubin:$PATH"

Linux:

• Ensure flock (util-linux) and timeout (coreutils) are available in PATH.

Installation details

Global install (one time)

scripts/setup-global.sh

This installs:

• ~/.codex/neural-codex/ (prompts, templates, skills, scripts, config stub)
• ~/.codex/prompts/ (so /prompts:neural.* appear)
• ~/.codex/skills/ (optional autodiscovery)

Use --force to overwrite existing files:

scripts/setup-global.sh --force

Project install (per repo)

scripts/setup-project.sh

This seeds a project with:

• .codex/prompts/
• .codex/templates/
• .codex/skills/
• .codex/config.toml (MCP stubs)
• scripts/neural-codex/ (loop + helpers)
• plans/prd.json, plans/progress.jsonl (from templates)

Install into another path:

scripts/setup-project.sh --path /path/to/project

Ralph loop usage

TEST_CMD="npm test" scripts/neural-codex/ralph-loop.sh 5

Notes:

• The loop claims one task per iteration from plans/prd.json.
• It writes progress to plans/progress.jsonl.
• It commits only when tests pass.

Memory workflow

• Use /prompts:neural.memory to append notes to plans/progress.jsonl.
• Use /prompts:neural.recall to search the log.
• For direct CLI usage: scripts/memory_write.py and scripts/memory_read.py.

Profiles

Named configuration sets for different workflows. Switch with codex --profile <name>:

Profile	Model	Approval	Use Case
default	gpt-5.2-codex	on-failure	Standard development
fast	gpt-4.1-mini	on-failure	Quick tasks, low cost
autonomous	gpt-5.2-codex	never	Ralph loop, unattended work
careful	gpt-5.2-codex	untrusted	Sensitive changes

Example:

codex --profile autonomous exec "Fix the auth bug"

MCP config

Supported MCP servers are stubbed in .codex/config.toml and include:

• chrome-devtools
• github
• search (Exa)
• optional playwright

Set tokens in your shell as needed (e.g., GITHUB_PERSONAL_ACCESS_TOKEN).

Advanced Config

The config file supports advanced options (see .codex/config.toml):

• Profiles: Named config sets with different models/approval policies
• Notifications: Webhooks, desktop alerts, CI integration
• History: Session transcripts with size caps
• Telemetry: OpenTelemetry for observability
• TUI: Clickable file citations (vscode, cursor, windsurf)

Reference: https://developers.openai.com/codex/config-advanced/

Repo layout

.
├── .codex/
│   ├── prompts/
│   ├── skills/
│   ├── templates/
│   └── config.toml
├── agents/
├── plans/
├── scripts/
└── README.md

Troubleshooting

Prompts not showing:

• Run scripts/setup-global.sh and restart Codex.

Ralph loop fails immediately:

• Ensure flock and timeout are in PATH.
• Ensure codex CLI is installed and logged in.

Tests not running:

• Set TEST_CMD explicitly for your project.

GitHub Pages

Static site lives in docs/. Enable Pages with source main / docs/.