Native Codex plugin · five focused skills

Evidence before confidence.

Discover the unknowns. Lock the plan. Build against it. Review independently. Exercise what the user will actually touch.

Install Neural Codex Read the workflow

One path. Five gates.

Each skill has a distinct authority and stopping condition. Small work can enter later; material work keeps planning, building, review, and behavior separate.

01 · map

$discover

Find architectural unknowns and produce an evidence-backed map.

02 · contract

$spec

Lock signatures, invariants, dependencies, and executable acceptance.

03 · build

$craft

Capture a baseline, implement the approved plan, and measure the delta.

04 · challenge

$vet

Review in fresh context and return an evidence-backed SHIP or HOLD.

05 · use

$exercise

Run tests, drive real behavior, and return PASS or FAIL.

Guardrails with visible limits.

Dependency-free Python hooks travel with the plugin. Codex does not trust them automatically: inspect them through /hooks first.

PreToolUse

Dangerous actions

Blocks high-confidence destructive shell commands and protected-branch force pushes.

PreToolUse

Sensitive files

Stops patch writes to common credential and environment files.

PostToolUse

Output scanning

Warns when tool output resembles secrets without echoing the detected value.

PreCompact

Context recovery

Preserves a small project note so work can recover after compaction.

Install from the repository marketplace.

No copy script and no silent configuration mutation. Add the marketplace, install the plugin, start a new task, then review its hooks.

$ codex plugin marketplace add brolag/neural-codex
$ codex plugin add neural-codex@neural-codex
$ codex plugin list

> /hooks
Five skills. One plugin. No shadow inventory.

The 1.0 structure removes unsupported command bundles, personas, profiles, templates, loops, and installers.

.agents/plugins/marketplace.json
plugins/neural-codex/
├── .codex-plugin/plugin.json
├── skills/
│   ├── discover/
│   ├── spec/
│   ├── craft/
│   ├── vet/
│   └── exercise/
└── hooks/
    ├── hooks.json
    └── *.py

Model choice stays with Codex and you.

Neural Codex never changes model, reasoning effort, approvals, sandboxing, network access, or global configuration. API availability does not guarantee Codex availability.

Available model first

Use the Codex default or a model your client explicitly exposes. Apply GPT-5.6 prompting principles without hardcoding an unsupported ID.

Permission stays separate

A stronger model is not broader authority. Choose approval and sandbox policy independently.

View configuration examplesReview the release gate