Build less. Prove it. Ship safely.
A lightweight software delivery plugin for Claude Code and OpenAI Codex, with one shared set of nine skills. It connects intent, planning, implementation, review, verification, release, and operations while sizing the process to the risk of the change.
The plugin adapts workflow ideas from Anthropic's AI-native SDLC playbook, Superpowers, GSD, BMAD, G-STACK, Ponytail, and Caveman. It uses original instructions; those frameworks are not dependencies or bundled runtimes.
Use a current Claude Code or Codex release with plugin support. The plugin itself needs no API keys, package installation, MCP servers, or background services. Your host still needs its normal account and tools.
Run in a terminal:
claude plugin marketplace add sofus-nl/ai-native-sdlc
claude plugin install ai-native-sdlc@ai-native-sdlc-marketplace --scope userStart a new Claude Code session, or run /reload-plugins in an existing one. Then:
/ai-native-sdlc:ai-native-sdlc Add CSV export to the existing reports page. Reuse the current data and UI components.
Run in a terminal:
codex plugin marketplace add sofus-nl/ai-native-sdlc --ref main
codex plugin add ai-native-sdlc@ai-native-sdlc-marketplaceStart a new Codex task so skill discovery reloads. Select ai-native-sdlc from the available skills, or ask:
Use the ai-native-sdlc plugin to add CSV export to the existing reports page. Reuse the current data and UI components.
Where workspace plugin import is available, a workspace admin can open Workspace settings > Plugins > Add > Import marketplace and enter:
- Source:
https://github.com/sofus-nl/ai-native-sdlc - Path: leave empty
- Branch:
main
Import the marketplace, then review the plugin's installation policy and workspace access. OpenAI accepts this repository's Claude-compatible marketplace format. See GitHub marketplace import and sync. Availability depends on your account, role, and product surface; workspace import has not been tested for this release.
Publishing this repository does not automatically list the plugin in a public directory. Uploading a ZIP into a regular conversation does not install it. Other agents may be able to load the skills/ files, but their compatibility is not verified here.
See the official Claude Code plugin documentation and OpenAI plugin packaging documentation. Use codex plugin --help for commands supported by your installed version.
Start with the router, ai-native-sdlc, when you want it to choose the next step. You can also request a specific skill:
| Skill | Purpose |
|---|---|
ai-native-sdlc |
Choose the risk lane and next action; resume existing work |
ai-native-shape |
Clarify the outcome, boundaries, and acceptance criteria |
ai-native-plan |
Plan concrete changes, dependencies, risks, and proof |
ai-native-build |
Implement the smallest change that meets the plan |
ai-native-debug |
Trace failures to their cause before changing code |
ai-native-review |
Check correctness, intent, risk, and unnecessary complexity |
ai-native-verify |
Gather fresh evidence for completion claims |
ai-native-ship |
Perform authorized Git delivery or a release, with its verification |
ai-native-operate |
Turn operational signals and incidents into follow-up work |
In Claude Code, prefix a skill with /ai-native-sdlc:, for example /ai-native-sdlc:ai-native-review. In Codex, select or name the skill.
Useful requests:
Use ai-native-debug to investigate the failing checkout test. Diagnose first.
Use ai-native-plan to plan this feature using existing components and no new dependencies unless required.
Use ai-native-review to review this diff against the acceptance criteria and identify unnecessary abstractions.
Use ai-native-verify to check the completed change and report what remains unverified.
Use ai-native-sdlc to resume the change recorded in .sdlc/changes/csv-export/state.md.
| Lane | When | Artifacts and checks |
|---|---|---|
| Fast | Small, clear, reversible changes without elevated risk | Inline plan, minimal edit, focused verification; no lifecycle files |
| Standard | Bounded features, ordinary bugs, and any change that adds a third-party dependency or changes the version of one | Specification, plan, state, review, and fresh verification |
| Controlled | Security, privacy, payments, migrations, production, destructive changes, or cross-service effects | Explicit intent and human gates, independent review, and recovery evidence |
Standard and Controlled work proceed through an ordered checkpoint list approved at the plan gate, each checkpoint passing its gates before the next; Controlled adds two context-isolated reviewers per checkpoint.
Standard and Controlled work keep durable artifacts under .sdlc/changes/<slug>/. Existing state identifies the next action, so work can resume without replaying the whole conversation. Only create the artifacts the chosen lane needs. When ai-native-ship finishes a change, it sets status: closed and moves the folder to .sdlc/archive/<slug>/, or removes it where repository instructions say so.
The agent follows repository instructions and the user's authorization. The skills do not grant permission to commit, publish, deploy, or change external systems. They are instructions, not a sandbox or an enforced policy engine.
The shared economy policy favors existing code, standard libraries, native capabilities, and the smallest sufficient change. It discourages speculative abstractions, duplicate artifacts, broad context loading, repeated reads, and continued implementation after acceptance passes.
Reference documents load only when needed. Concise output must preserve exact commands, identifiers, evidence, and important qualifications. No percentage token savings is claimed: savings require paired task runs measuring provider totals and result quality.
Delegation is conditional on task independence, context needs, and review requirements, subject to host, repository, and user instructions; a user's request for parallel or delegated work counts when the host supports it. Workers receive bounded scopes and return compact evidence; one owner integrates results and verifies the current change. Fresh, forked, or resumed context depends on host capabilities. See the reviewed rationale.
Model selection uses an on-demand policy: retain the configured baseline until task-specific evidence supports a cheaper model/effort configuration. The eight named model families are candidates, not preset tiers. Native host controls govern execution; unsupported overrides and unverified model identity are reported. Complete-task cost includes retries and verification. No automatic router, cross-provider access, or guarantee of equal quality is bundled. See the reviewed rationale.
For the software you build with this plugin, planning, verification, and shipping use the SemVer reference. Agents identify the public contract, justify the version bump against the complete release, check version consistency, and preserve released versions. Explicit repository versioning policies are not silently migrated. Ordinary non-release edits do not trigger automatic bumps or publication.
A plugin runs with your rights, so review it before you trust it. This one ships Markdown skills and, on Claude Code, one hook: hooks/edit_guard.py. Claude Code runs a plugin hook with your full user permissions, outside the Bash sandbox. The hook reads the session transcript and .sdlc/ files, hashes the project's files to see what changed, keeps a small record per session in the plugin's data folder, and makes no network connection; SECURITY.md has the full list. On Codex the plugin registers no hook. To list what an installed plugin registers on Claude Code, run claude plugin details ai-native-sdlc.
An update replaces the files you reviewed. To stay on a reviewed revision, pin the marketplace when you add it:
claude plugin marketplace add sofus-nl/ai-native-sdlc@v0.10.0
codex plugin marketplace add sofus-nl/ai-native-sdlc --ref v0.10.0-
Claude Code, tag: the suffix takes a branch or a tag. Auto-update is off by default for a third-party marketplace such as this one; check it under
/plugin> Marketplaces. -
Claude Code, commit: from Claude Code 2.1.324 the marketplace
refalso takes the full lowercase SHA of a commit that is on a branch or tag, and a refresh cannot move it. The documentation gives this as a settings source object and shows no command-line form. In.claude/settings.json, with the commit you reviewed (here the one behindv0.9.0):{ "extraKnownMarketplaces": { "ai-native-sdlc-marketplace": { "source": { "source": "github", "repo": "sofus-nl/ai-native-sdlc", "ref": "5f69bf2aa1139565d4b2ae8f88779837a5e6b086" } } } }Start a Claude Code session in that project once, so the marketplace loads, then run
claude plugin install ai-native-sdlc@ai-native-sdlc-marketplace --scope project. See the marketplace reference and plugin security. -
Codex:
--refpins a Git ref. The Codex plugin documentation does not say whether it takes a commit SHA.
This project never moves a release tag (see versioning). git ls-remote --tags https://github.com/sofus-nl/ai-native-sdlc lists the tags; the line that ends in ^{} shows the commit behind a tag.
To report a vulnerability, see SECURITY.md.
- Skills are missing: verify installation with
claude plugin list --jsonorcodex plugin list, then start a fresh session/task. - You installed an earlier local version: it may coexist under
ai-native-sdlc@personalin Codex. Keep one active copy if duplicate skills appear. - Too much ceremony: state the actual scope and ask the router to explain its lane choice. A genuinely low-risk edit should use Fast.
- Controlled gates skipped on
claude-sonnet-5-5: in 0.5.0 behavior checks, this model did not reliably follow Controlled gates such as the separate test author, final verification, and the specification's Reference oracle. It is not yet qualified for Controlled work. Useclaude-sonnet-5or another qualified model, and confirm thatstate.mdrecords intent, design, and plan approvals before any code changes. - Writes denied with "State the lane" or "records these approvals": in Claude Code, the plugin's guard hook (
hooks/edit_guard.py, needspythonon PATH) blocks file changes until the lane is stated and the active change'sstate.md(the one the session last wrote, or else last read) records the required approvals. Ask the router to state the lane, or record the approval you gave. A command that adds a package (npm install <package>,pip install <package>, and the like) counts as such a file change: it needs a stated lane, and in Standard and Controlled work the approvals. A restore such asnpm cidoes not count. The guard is not available on Codex. - A probe, benchmark, or other throwaway file denied before the plan is approved: such files are open in two places. One is the system temp folder. The other is any file that Git ignores, so a scratch folder inside the project needs one line in
.gitignore(for example.scratch/). They do not belong under.sdlc/. - A
state.mdwrite denied with "records status: verified only after": in Standard and Controlled work, the guard lets the agent recordstatus: verifiedonly after it invoked the verify skill in the same session, on the files its last review saw. The message names the skill to invoke and the files that changed since the review; if one of those is test output and not code, remove it or let Git ignore it. If you verified the change yourself, editstate.mdby hand. - Writes denied with "guard failed", or a hook error that says "guard is inactive": the guard hook itself hit an error. "Guard failed" denies the write because the session uses the plugin and the approvals could not be checked; "guard is inactive" blocks nothing. Both name the cause. Report it in an issue; to continue meanwhile, disable the plugin.
- Other workflow plugins compete: invoke this plugin explicitly and disable overlapping plugins if their instructions conflict.
The repository root is the plugin root. .claude-plugin/ supplies the Claude manifest and marketplace; .codex-plugin/ supplies Codex metadata; both hosts use the same physical skills/ tree.
git clone https://github.com/sofus-nl/ai-native-sdlc.git
cd ai-native-sdlc
claude plugin validate .claude-plugin/plugin.json --strict
claude plugin validate .claude-plugin/marketplace.json --strictKeep changes focused, preserve cross-host compatibility, and check relative references and skill discovery in both hosts. Follow Semantic Versioning 2.0.0 using the public API and release policy. Run python scripts/check_version.py to check version syntax, manifest alignment, and the current changelog entry; CI runs the same check. Record changes in the changelog. Open an issue with a concrete example or a pull request with the problem, change, and verification evidence. Do not include secrets or private project data.
Before releasing skill changes, run the behavior checks for risk routing, reuse, authorization, diagnosis, stopping, and delegation. On Claude Code, claude plugin eval runs 20 of the 22 behavior cases (those in evals/) with one command. The fixture generator and filesystem checker use Python's standard library; transcript review checks the agent's actual actions.
MIT. See third-party method references for sources and attribution. This project is not affiliated with or endorsed by Anthropic, OpenAI, or the referenced framework authors.