codex-workflows

codex-workflows

On larger product work, Codex can pursue technical consistency beyond what the user needs. Handling every edge case and making each path deterministic can alter what users see even when the approved outcome does not require it.

codex-workflows keeps that work within the smallest approved outcome. It checks scope and rough cost before design, records user-visible contracts and exclusions, and creates only the design and verification work the change needs. Codex still chooses the implementation details that stay within those boundaries.

The workflows are installed as Agent Skills and custom agents for OpenAI Codex CLI. The main Codex session owns progress, evaluates specialist findings against the approved outcome, and resolves implementation details from repository evidence.

Why not use Codex directly?

Direct Codex is the better fit for a well-scoped fix, disposable experiment, or one-shot script. It is faster and cheaper when the intended outcome and safe implementation boundary are already clear.

Use codex-workflows when technical choices can change the product scope, user-visible behavior, or a decision that needs to survive across contexts.

For example, a request to extend an existing authentication path can lead to a technically cleaner second mechanism, broader validation, and a new response contract. The frontend may adapt and the tests may pass, while users receive behavior that was never part of the approved change.

codex-workflows controls that expansion at three points:

This workflow uses more agent calls and tokens than direct execution. Use it when protecting the approved outcome is worth that cost.

An edge case does not require work simply because Codex can handle it. Additional validation, deterministic behavior, or a new abstraction must protect an approved requirement, an observable contract, or a demonstrated failure.

A real workflow run

The BytePlus Seedream provider integration in mcp-image added a third external image provider across 18 files. Eight planned tasks kept the public MCP request, client, file-save, and file-URI contracts unchanged while the provider-specific implementation evolved.

Before merge, live evaluation established the final model routing, prompt limits, timeout, and response handling. Independent reviews also caught an unbounded file read, a validation bypass, a blocking FIFO path, and inconsistent API-key normalization. All four were fixed, and the PR passed 303 tests across 19 files plus a no-retry live provider call. Across the eight tasks and four fixes, the approved public contracts stayed unchanged.

Quick Start

Requires Node.js 22 or later and the latest Codex CLI.

Install and run

cd your-project
npx codex-workflows install

Then invoke a recipe in Codex CLI:

$recipe-implement Add user authentication with JWT

$ invokes a skill explicitly. Type $recipe- to see the available workflows.

Choose a path

How It Works

flowchart LR
    A[Request] --> B[Agree on the smallest useful outcome]
    B --> C{Design needed?}
    C -->|No| H[Execute autonomously]
    C -->|Yes| D[Inspect and record needed decisions]
    D --> E[Approve product and major design boundaries]
    E --> F[Plan dependent work]
    F --> G[Approve the implementation scope]
    G --> H
    H --> I[Verify approved outcome]
    I -->|Fixable implementation gap| H
    I -->|Requirement or major design changed| B
    I -->|Passed| J[Complete]

The number of independent product and design decisions determines the route. File count and the number of edge cases Codex can identify do not:

An ADR is created only for a durable current-scope choice with at least two materially distinct options. When several choices qualify, their ADRs are reviewed together. An integration or E2E test is selected only when a cheaper test cannot prove the required interaction. Some changes need neither.

The documents record only decisions that affect the product or repository implementation. Third-party approval, production access, release execution, and unrelated operational work do not become implementation gates.

After the implementation scope is approved, the orchestrator runs the tasks, focused verification, applicable repository checks, and one implementation commit per task. It resolves problems from the approved documents and repository evidence first. User-visible behavior remains a product boundary rather than something the implementation may adjust for internal consistency. The orchestrator asks you only when progress requires a new product requirement, a change to a major approved design decision, authority only you hold, or an irreversible action you did not authorize.

Specialist agents receive the exact documents and paths needed for their work. They supply focused evidence without inheriting authority to expand the approved outcome.

A handoff you can inspect

The included Work Plan template ties each implementation task to its Design Doc section and acceptance criteria:

### P1-T1: Preserve the error response contract

- **Source**: `docs/design/example-design.md`, API contract, AC-2
- **Scope**: Update the repository implementation and its focused tests
- **Depends on**: none
- **Verification**: Run the contract test and observe the documented response shape

The Task File Contract carries the source, intended result, target files, and executable verification into implementation. It adds a Verification Focus only when a test could pass without proving one important behavior. Final review checks the approved documents against the completed diff.

Installation

Requirements

• Codex CLI (latest)
• Node.js >= 22

Install

Install into the current project:

cd your-project
npx codex-workflows install

This copies into your project:

• .agents/skills/: Codex skills (foundational + recipes)
• .codex/agents/: Subagent TOML definitions
• Manifest file for tracking managed files

To make the workflows available to Codex across all projects, install them into your user-level CODEX_HOME instead:

npx codex-workflows install --user

This installs skills into $CODEX_HOME/skills/ and agents into $CODEX_HOME/agents/. When CODEX_HOME is not set, it defaults to ~/.codex.

Update

# Preview what will change
npx codex-workflows update --dry-run

# Apply updates
npx codex-workflows update

# Update a user-level installation
npx codex-workflows update --user

The updater preserves files you have modified locally. It compares each file against its hash at install time and skips changed files. Versioned update history applies file moves and deletions in order, so local changes follow a moved file to its current path. Modified files retired without a replacement are moved to .codex-workflows-preserved/<version>/. New files from the update are added automatically.

# Check installed version
npx codex-workflows status

# Check a user-level installation
npx codex-workflows status --user

Workflow Recipe Reference

Invoke recipes with $recipe-name in Codex. Type $recipe- and use tab completion to see all available recipes.

Working State

Recipes use docs/plans/ as ephemeral working state for Work Plans, implementation Task Files, and temporary review-fix or test-addition Task Files. Task and phase progress is updated there after each quality-approved implementation commit, while those progress files stay outside that commit. Add the directory to your project's .gitignore unless your team intentionally wants to review those transient files:

docs/plans/

PRDs, ADRs, UI Specs, and Design Docs are durable project documents and are intended to be committed.

Included Guidance

Recipes load the repository-aware guidance required for the current task. You rarely need to select these skills directly.

Specialized Agents

Codex spawns these as needed during recipe execution. You do not need to learn them first; recipes route domain work to the relevant agents while the orchestrator retains workflow control. Each agent runs in its own context with specialized instructions and explicitly named required skills.

Project Structure

After installation, your project gets:

your-project/
├── .agents/skills/           # Codex skills
│   ├── coding-rules/         # Foundational guidance
│   ├── testing/
│   ├── ai-development-guide/
│   ├── documentation-criteria/
│   ├── requirement-convergence/
│   ├── implementation-approach/
│   ├── integration-e2e-testing/
│   ├── external-resource-context/
│   ├── llm-friendly-context/
│   ├── task-analyzer/
│   ├── subagents-orchestration-guide/
│   └── recipe-*/             # Workflow entry points ($recipe-*)
├── .codex/agents/            # Subagent TOML definitions
│   ├── requirement-analyzer.toml
│   ├── technical-designer.toml
│   ├── ui-analyzer.toml
│   ├── task-executor.toml
│   └── ... (25 agents total)
└── docs/                     # Created as you use the recipes
    ├── prd/
    ├── design/
    ├── adr/
    ├── ui-spec/
    └── plans/
        └── tasks/

Works With

If your requirements already live in Linear or an existing PRD, linear-prism can decompose them into implementation-ready tasks by reading the codebase, making dependencies explicit, and preserving Design Doc boundaries.

Those tasks can then be passed into $recipe-design to enter the design phase with clearer scope and better task visibility.

FAQ

Q: What models does this work with?

A: Designed for current GPT models. Models are configurable per agent in the TOML files.

Q: Can I customize the agents?

A: Yes. Edit the TOML files in .codex/agents/ to change model, sandbox_mode, or developer_instructions. Each agent names its required skills in developer_instructions. Files you modify locally are preserved during npx codex-workflows update.

For a user-level installation, edit the files in $CODEX_HOME/agents/ and use npx codex-workflows update --user. User-level files modified after installation are preserved in the same way.

Q: What's the difference between $recipe-implement and $recipe-fullstack-implement?

A: $recipe-implement is the universal entry point. It runs requirement-analyzer first, uses the request and repository scope to identify affected layers, and automatically routes to backend, frontend, or fullstack flow. $recipe-fullstack-implement skips the detection and goes straight into the fullstack flow (separate Design Docs per layer, design-sync, layer-aware task execution). Use $recipe-implement when you're not sure; use $recipe-fullstack-implement when you know upfront that the feature spans both layers.

Q: Does this work with MCP servers?

A: Yes. Codex skills and subagents work alongside MCP. Skills operate at the instruction layer, while MCP operates at the tool transport layer. Custom agents inherit parent mcp_servers when the agent TOML omits mcp_servers; add agent-local MCP config only for agent-specific servers or tool filtering.

Q: How is this related to claude-code-workflows?

A: claude-code-workflows is the Claude Code counterpart. The repositories share the same workflow philosophy, adapted to each tool's native extension points. They can coexist in the same project because codex-workflows installs its agent definitions under .codex/agents/ and Claude Code uses its own .claude/ files.

Q: What if a subagent seems stuck?

A: The main Codex session owns progress. It inspects the returned evidence, retries or repairs unusable results, and continues unaffected work. A subagent result does not stop the workflow by itself.

Design Rationale

License

MIT License. Free to use, modify, and distribute.

Built and maintained by @shinpr