← Kai’s cornerA project by Kai

Your repos. One coordinated delivery.

One request.
An AI team
across your repos.

Connect your requirements, backend, and frontend repositories. Describe a feature in Codex. codex-sdlc coordinates the AI roles, checks their handoffs, and brings the delivery back for your review.

Map your repositories once. Start each feature with an outcome.

Open source · Apache 2.0 · Guide: 0.5.0

YOU / THE REQUEST

“Let signed-in users save favorite items and find them later.”

AI · PMOne plan. Checked handoffs.
REPO 01 · COORDINATORRequirements/work/requirementsAI · BA

Requirements &
shared run record

REPO 02Backend/work/apiAI · BE

API, storage &
permissions

REPO 03Frontend/work/webAI · FE

Interface &
user experience

Reviewed API before BE + FE implementation
PM · IntegrationIndependent QC
YOU / THE DECISIONReview the evidence. Accept or request changes.
◆ You guide & accept ◆ AI coordinates & delivers Illustrative repository map

SDLC means Software Development Life Cycle.

One structured, resumable workflow. Across the repositories your product already uses.

Meet codex-sdlc

Watch the idea unfold.

Start with the introduction, then follow a feature across three repositories in the interactive walkthrough.

Explore the workflow
codex-sdlc · Introduction Watch on YouTube

Follow one feature across three repos

The handoffs.
Handled.

You describe the outcome. The PM carries context between specialists, reviews their work, and keeps one delivery record.

Illustrative walkthroughNo agents or repository changes are run on this page.

01 · REQUIREMENTS / COORDINATOR/work/requirementsShared delivery record
02 · BACKEND/work/apiAPI & backend work
03 · FRONTEND/work/webWeb interface work

01 / YOU → AI PM

Start with the outcome.

In your initialized workspace, tell Codex what you want. The PM inspects the mapped repositories and plans the work in the requirements repo.

Input
Your feature request + repository mappings
Output
One recorded run and scoped assignments
THE FEATURE REQUEST
“Let signed-in users save favorite items and find them later.”

No need to relay the same request to each role.

02 / AI BA + PM · REQUIREMENTS REPO

Make the behavior testable.

The BA turns the request into requirements, edge cases, and acceptance criteria. The PM reviews them before handing work to implementation roles.

Input
Request + repository context
Output
Reviewed requirements in the coordinator
EXAMPLE ACCEPTANCE CRITERIA
  • AC-01 Signed-in users can save an item.
  • AC-02 Favorites persist across sessions.
  • AC-03 A user's list is private to them.
Where does BA work live?

In the requirements repository, under .sdlc/runs/<run-id>/artifacts/ba/. This repository is the coordinator in our example.

03 / AI BACKEND + PM · API REVIEW

Agree on the connection.

Backend defines the API against reviewed requirements. Frontend implementation waits for contract review, so both sides build from the same agreed interface.

Input
Reviewed requirements
Output
Reviewed API contract for backend + frontend
EXAMPLE CONTRACT
  • PUT/favorites/{itemId}
  • GET/favorites
  • AUTH Return only the signed-in user's items.

◇ Gate: PM reviews the contract before implementation starts.

04 / AI BACKEND + FRONTEND · SEPARATE REPOS

Two codebases. One agreement.

Each role receives its repository, permitted paths, and reviewed inputs. Ready tasks can run in parallel when host capacity and safe scopes allow.

Input
Reviewed contract + bounded assignments
Output
Per-repository changes and implementation handoffs
ASSIGNED WORK
AI · BE
Backend repository

Favorites storage, API behavior, and access controls.

AI · FE
Frontend repository

Save control, favorites view, loading and error states.

05 / AI PM · ACROSS THE MAPPED REPOS

Check the handoffs. Connect the work.

The PM reconciles reported changes with assignments and reviewed inputs. Integration checks run against the mapped repositories and record where each command ran.

Input
Backend + frontend handoffs
Output
Integrated work and verification records for QC
Inspect the handoff chain

Requirements revision → reviewed contract → repository and assigned paths → reported changes → verification record. Inconsistent or stale handoffs need correction before acceptance.

A HANDOFF CAN BE BLOCKED

An old contract should not silently become the new truth.

Show a blocked handoff
Paused · Frontend handoff uses an older contract.

PM returns it to frontend with the current contract and assignment. The frontend role reconciles its work, reruns the required checks, and resubmits. Backend work remains recorded.

PM → Frontend → Recheck → Integration

06 / AI QC · INDEPENDENT VERIFICATION

Find the gaps. Route the repair.

QC checks the integrated feature against acceptance criteria, permissions, negative cases, and regressions. A defect goes back to the responsible role through the PM.

Input
Integrated feature + acceptance criteria
Output
QC findings, evidence, and any required rework
EXAMPLE DEFECT LOOP

Can one user read another user's favorites?

Show QC rework
Defect · Backend permission check needs repair.
  1. QC records the failing case and evidence.
  2. PM assigns the repair to backend in /work/api.
  3. Backend fixes and hands off the change.
  4. Integration checks and independent QC run again.
QC → PM → Backend → Integration → QC

Retest evidence is required before this defect can close.

07 / AI PM → YOU

The final decision stays yours.

When the required reviews and checks are complete, the PM assembles the delivery. You see what changed across repositories, what was verified, and any limitations before accepting or requesting changes.

Input
Reviewed handoffs + independent QC evidence
Output
Delivery package for human acceptance
YOUR DELIVERY PACKAGE
  • Changes, grouped by repository
  • Acceptance coverage and command evidence
  • Decisions, limitations, and remaining blockers

You accept the delivery or request changes.
Production deployment is a separate decision.

Step 1 of 7

Coordination you can inspect

Separate repos.
Clear ownership.
Checked handoffs.

The framework makes the boundaries explicit: where a role works, which inputs it uses, and what evidence it returns.

01 / LOCATION

Confirm the checkout.

Mappings check the Git remote and checkout root. Missing, mismatched, duplicate, or nested mappings are rejected.

Right repository, before the assignment.
02 / OWNERSHIP

Bound the assignment.

Tasks carry repository IDs and permitted paths. Handoff validation checks reported changes against ownership and protected-path rules.

Backend work has a backend scope.
03 / DEPENDENCIES

Share reviewed inputs.

Frontend waits for API review. Stale or inconsistent assignments and reports must be corrected before their handoffs are accepted.

A handoff is a checked dependency.
04 / CONTINUITY

Keep one delivery record.

Per-run locks and version checks protect published run records. Evidence identifies the working repository, and saved state supports resuming.

The next session can pick up the record.

These checks help keep changes within scope and surface invalid handoffs. They do not guarantee conflict-free work or replace your review.

An AI team with defined responsibilities

Specialists do the work.
A PM connects it.

Roles are responsibilities, not repositories. The PM and QC follow the delivery across every mapped checkout.

PM

Coordinate the delivery

Dispatch ready work, review requirements and contracts, reconcile handoffs, and assemble the result.

BA

Define the behavior

Turn the request into traceable requirements and acceptance criteria in the coordinator.

BE

Build the backend

Own the assigned API, storage, and backend implementation in its mapped repository.

FE

Build the experience

Implement web or mobile interfaces against reviewed contracts in the assigned repository.

QC

Verify independently

Check the integrated feature, report defects, and retest the repairs.

PO · OPTIONAL

Advise on the outcome

The AI Product Owner reviews business value and acceptance coverage. Human acceptance stays with you.

Choose models by role

Start with inherited Codex settings, or use setup to configure a model and reasoning effort for each role. For example, you can request different models for frontend, backend, and QC.

These are configuration preferences; availability and dispatch depend on the Codex host. Read the model configuration guide ↗

Map once. Deliver feature by feature.

Bring your repos.
Start with an outcome.

Use your requirements or BA repository as the coordinator. Keep backend and frontend source in their own repositories, with their own Git histories.

Before setup
  • Node.js >=24.16.0 <25 and npm 11
  • Existing application checkouts accessible from your Codex session
  • Each mapped checkout is an independent Git root with an origin remote
  • Verification commands and required services available for checks
Where everything lives

The coordinator's .sdlc/ keeps the run record, BA artifacts, handoffs, and evidence. Machine-specific paths stay in ignored .sdlc/local.yaml.

Explore multi-repository setup ↗
  1. 01

    Install for Codex

    Add codex-sdlc from the Plugins Directory.

    Open plugin page
  2. 02

    Connect your three repositories

    Open the requirements checkout in Codex. Replace the path placeholders and send:

    MULTI-REPOSITORY SETUP
    Set up codex-sdlc in multi-repository mode. Use this requirements repository as the coordinator. My backend checkout is at <absolute backend path>, and my frontend checkout is at <absolute frontend path>.
    Inspect the applications, confirm the repository mappings and verification commands, and show the setup dry run before applying changes.
    One repository? Start here instead.

    Open your existing application in Codex and use:

    Initialize codex-sdlc for this existing project. Explain the setup choices and show the dry run before applying changes.
  3. 03

    Give the team a feature

    After setup checks pass, send one outcome. Adapt this example to your product:

    YOUR FIRST FEATURE
    Start a codex-sdlc feature delivery across the mapped repositories: Let signed-in users save favorite items and find them later.
    Coordinate BA, backend, frontend, integration, and independent QC. Bring the result back for my acceptance.

Prefer direct commands? Read the complete setup guide.

Before you start

The useful questions.

What does the plugin add to Codex?

It provides delivery skills for setup, PM coordination, business analysis, backend and frontend implementation, independent QC, and optional Product Owner advice. The repository runtime supplies the workflows, policies, schemas, and recorded state.

Do I need an existing application?

Yes. Initialization configures existing application directories; it does not scaffold a complete product from an empty folder. Built-in presets cover Go, Next.js, Flutter, PostgreSQL, and Redis, with generic application support for other stacks.

Why are the skills not in my project's .agents folder?

When installed from the Plugins Directory, the skills live in the Codex plugin cache. Your project receives .sdlc/ and a managed AGENTS.md block, so a project-level skills folder is not required.

What happens if the workflow pauses halfway?

The shared run record keeps assignments, evidence, decisions, and blockers. Codex asks for unclear requirements, material decisions, or approvals required by your project policy, then uses recorded state to resume. Partial work remains visible; separate repositories are not rolled back as one atomic transaction.

Can my BA repository be separate from the coordinator?

In this example, the BA repository is the coordinator. BA outputs belong under its .sdlc/runs/<run-id>/artifacts/ba/ directory. Mapping a separate documentation repository does not grant BA automatic write access to arbitrary documents there.

Does updating the plugin update my project?

No. The plugin update supplies current skills. An initialized project's pinned runtime is upgraded separately, with a dry run and rollback support. Follow the upgrade guide.

Where does delivery information live?

Configuration, requests, run manifests, reports, decisions, defects, and evidence remain in the coordinator repository under .sdlc/. Read the project's privacy guidance for details.

One request starts the work

Your next feature.
Across the repos you already have.