Version 1.0.0, released on July 15, 2026.
This kit adds project-local programming agents to software repositories without exposing them in repositories that contain your agent tools.
The agents work across technology stacks, including:
- SvelteKit
- Flutter and Dart
- Python
- repositories that combine these stacks
- other stacks with documented project conventions
Install the core agents in a target repository before reviewing the optional files.
| Source in this kit | Destination | Required | Purpose |
|---|---|---|---|
project/.codex/ |
<target-repository>/.codex/ |
Yes | Agent definitions and project agent settings |
project/AGENTS.md |
<target-repository>/AGENTS.md |
Yes | Agent selection and coordination rules |
project/optional-agents/*.toml |
<target-repository>/.codex/agents/ |
No | Deep exploration or Git publishing |
global/AGENTS.md |
~/.codex/AGENTS.md |
No | Personal development instructions across repositories |
global/config.toml.reference |
~/.codex/config.toml |
No | Reference values for personal Codex configuration |
Do not place the project/ directory itself inside the target repository. Copy its .codex/ directory and AGENTS.md file to the repository root.
Run these commands from the target repository root. Replace the first path with the location of this kit:
KIT=/path/to/codex-dev-agents-kit
mkdir -p .codex/agents
cp -R "$KIT/project/.codex/." .codex/The target repository now contains:
<target-repository>/
|-- .codex/
| |-- config.toml
| `-- agents/
| |-- code_explorer.toml
| |-- quick_implementer.toml
| |-- implementer.toml
| `-- code_reviewer.toml
`-- ...
If the target repository has no AGENTS.md:
cp "$KIT/project/AGENTS.md" ./AGENTS.mdIf it already has AGENTS.md, copy the Project-local programming agents section from project/AGENTS.md into the existing file. Keep the repository's current instructions.
Commit .codex/ and AGENTS.md when everyone who uses the repository should receive the agents.
For a local-only installation, add these entries to the target repository's .gitignore before staging the files:
# Local Codex agent configuration
/.codex/
/AGENTS.mdIgnore /AGENTS.md only when the file exists solely for this kit. Do not ignore a shared AGENTS.md that the repository already tracks. Git does not apply new ignore rules to tracked files; remove them from the Git index yourself if you choose to convert an existing installation to local-only.
Start a Codex session from the target repository and test read-only exploration:
Use the agents to analyze this repository's structure. Do not modify files.
The parent agent should select code_explorer. Use /agent to inspect agent activity when testing the setup.
~/.codex/ # Optional personal configuration
|-- AGENTS.md
`-- config.toml
workspace/
|-- target-project-a/
| |-- AGENTS.md
| `-- .codex/
| |-- config.toml
| `-- agents/
|-- target-project-b/
| |-- AGENTS.md
| `-- .codex/
| |-- config.toml
| `-- agents/
|-- agent-tool-a/ # No project installation
`-- agent-tool-b/ # No project installation
Install the project files in software repositories that need the agents. Do not install them in repositories that contain agent tools.
The files under <repository>/.codex/agents/ define custom Codex subagents. The repository-root AGENTS.md tells the parent agent when to use them.
- Parent agent: the Codex agent that receives the user's request and owns requirements, decisions, integration, verification, and the final response.
- Subagent: a custom agent that the parent agent starts for a bounded task.
- Project-local agent: a subagent defined under one repository's
.codex/agents/directory. - Working-tree review: a review of staged, unstaged, and relevant untracked changes.
- Comparison review: a review of committed work against a branch, tag, or commit.
- Commit message: a Conventional Commits-compliant description of a Git commit. A request for a message does not authorize staging or committing.
| Agent | Model | Reasoning effort | Sandbox mode | Availability |
|---|---|---|---|---|
code_explorer |
gpt-5.6-luna |
high |
read-only |
Core |
quick_implementer |
gpt-5.6-luna |
high |
workspace-write |
Core |
implementer |
gpt-5.6-sol |
low |
workspace-write |
Core |
code_reviewer |
gpt-5.6-sol |
low |
read-only |
Core |
deep_code_explorer |
gpt-5.6-luna |
xhigh |
read-only |
Optional |
commit_publisher |
gpt-5.6-luna |
high |
workspace-write |
Optional |
This release targets OpenAI's current GPT-5.6 Codex model family: Sol, Terra, and Luna.
- The implementation and review roles use
gpt-5.6-sol. - The exploration, focused-change, and Git publishing roles use
gpt-5.6-luna. - The supplied agents do not select
gpt-5.6-terra. Users who prefer Terra for an everyday role can change that agent'smodelfield.
The kit does not pin the parent model. Select a model and reasoning effort in the Codex CLI session through /model.
This kit does not require or invoke the external $code-review skill.
The code_reviewer prompt incorporates the relevant parts of that workflow:
- working-tree, comparison, combined, and supplied-diff review modes
- validation of the Git comparison target and review scope
- requirement checks when a reliable specification exists
- checks against documented repository standards
- separation of confirmed defects, requirement mismatches, documented-standard violations, heuristic concerns, and unresolved questions
- project rules taking precedence over generic heuristics
- checks for missing requirements, partial implementations, incorrect behavior, regressions, and unnecessary scope
The reviewer remains one read-only subagent and does not start more reviewers. Its report covers Correctness and Risk, Spec, Standards, Tests, and UI when the change affects visible UI.
You can uninstall $code-review without losing the checks listed above.
Install $devibify separately when you want the agents to apply its UI workflow. Use it for work that changes visible UI:
quick_implementeruses it for small UI changes.implementeruses it for larger UI work and its final quality pass.code_revieweruses it as a read-only diagnostic guide.- The parent agent uses it for UI planning, implementation, or review.
The kit uses the existing implementation agents for UI work instead of adding a separate UI agent.
Apply UI guidance in this order:
- The user's requirements
- The project's existing design system and conventions
- Project-specific UI instructions
$devibifydefaults and checklists
codex-dev-agents-kit/
|-- .gitignore
|-- README.md
|-- CHANGELOG.md
|-- SOURCES.md
|-- THIRD-PARTY-NOTICES.md
|-- global/
| |-- AGENTS.md
| `-- config.toml.reference
|-- project/
| |-- AGENTS.md
| |-- .codex/
| | |-- config.toml
| | `-- agents/
| | |-- code_explorer.toml
| | |-- quick_implementer.toml
| | |-- implementer.toml
| | `-- code_reviewer.toml
| `-- optional-agents/
| |-- deep_code_explorer.toml
| `-- commit_publisher.toml
|-- examples/
| `-- OPTIONAL-PROJECT-GUIDANCE.md
`-- guides/
|-- ADOPTION.md
|-- PROMPTS.md
`-- WORKFLOWS.md
The core agents work without the files under global/.
Compare global/AGENTS.md with your existing ~/.codex/AGENTS.md and merge the rules you want to apply across repositories. Do not overwrite personal instructions without reviewing the differences.
global/config.toml.reference contains reference values for the MultiAgent V2 feature block. Merge values that your Codex CLI version still requires into ~/.codex/config.toml. The reference is not a complete personal configuration and does not set the parent model.
Do not copy global/AGENTS.md into a target repository. Do not copy project/AGENTS.md into ~/.codex/.
Set KIT to this kit's path if you opened a new shell:
Install the deep explorer:
KIT=/path/to/codex-dev-agents-kit
cp "$KIT/project/optional-agents/deep_code_explorer.toml" .codex/agents/Install the Git publishing agent:
cp "$KIT/project/optional-agents/commit_publisher.toml" .codex/agents/The commit_publisher performs Git mutations. Leave it uninstalled when you want to control staging, commits, and pushes yourself.
Do not copy project/AGENTS.md or project/.codex/agents/ into repositories that contain your agent tools.
You do not need to specify a model, reasoning effort, sandbox mode, agent_type, fork_turns, or thread parameters.
Use the agents when they help, and implement this change: <description>.
Use code_explorer to trace the flow, then fix the problem.
Use code_reviewer to review all current changes. Fix concrete problems and verify the result.
Use code_reviewer to compare this work with main and check it against the requirements I provided.
Use the agents and $devibify to implement and verify this UI.
Do not use subagents for this change. Handle it in the parent agent.
See guides/PROMPTS.md and guides/WORKFLOWS.md for more examples.
The supplied project/AGENTS.md contains agent selection and coordination rules. Add project details when they remain stable and are hard to infer, such as:
- exact commands that project files do not reveal
- generated directories that agents must not edit
- architectural boundaries
- separate toolchains in a mixed repository
- external service setup
- platform or deployment constraints
See examples/OPTIONAL-PROJECT-GUIDANCE.md.
Parse every supplied TOML file and the global reference:
python3 - <<'PY'
from pathlib import Path
import tomllib
files = sorted(Path("project").rglob("*.toml"))
files.append(Path("global/config.toml.reference"))
for path in files:
tomllib.loads(path.read_text())
print("Parsed %d configuration files." % len(files))
PYFor a runtime check, start a new Codex session in a trusted repository and run:
Use the agents to analyze this repository's structure. Do not modify files.
Use /agent to inspect the initial test run or troubleshoot agent routing.