Governed Agent SDLC

Ship with agents. Keep humans in control.

Project website · Architecture · Workflow · CLI reference · Adoption guide

Governed Agent SDLC keeps one workflow policy for four AI coding agents: Claude Code, Codex, OpenCode and Antigravity. The roles, artifact lifecycle, approval rules and protected areas live in a tool-neutral core and one project manifest, and agentkit generate turns them into each agent's own configuration; none of them keeps a parallel policy file. Stack profiles supply each stack's commands and exclusions. Maturity says which guards each agent enforces.

Human approvals can be checked against GitHub. With [review].approval_mode = "github" (or --approval-mode github on an artifact transition), agentkit artifact transition ... approved and agentkit review approve accept an approval only when its evidence is a pull request review (or, for an artifact, an issue or pull request comment) in this repository, written by a human owner, member or collaborator, as looked up through gh. In the default local mode an approval is recorded as given and marked unverified.

The project is intentionally human-in-the-loop. It helps agents work predictably; it does not grant an AI permission to approve, merge, release, or retrieve credentials.

Alpha status: the core workflow, the Claude Code, Codex, OpenCode and Antigravity adapters, deterministic safety hooks, cross-platform validation, and PyPI Trusted Publishing are available. Interfaces may still evolve before 1.0.

What is included

• Six tool-neutral roles: architect, planner, developer, reviewer, QA, and publisher.
• A structured artifact lifecycle with explicit approval evidence and supersession.
• A TOML project manifest for repository layout, profiles, commands, and protected areas.
• A dependency-free Python CLI: init, generate, update-hooks, upgrade, validate, doctor, and artifact commands.
• Claude Code agent generation and safety hooks.
• Native Codex project configuration and governed role generation.
• OpenCode subagents with a guard plugin, and Antigravity agents with a guard hook (OpenCode adapter, Antigravity adapter).
• Stack profiles for generic repositories, Python, .NET, Nuxt, Node.js and TypeScript, Java (Spring Boot on Maven), Flutter, Go, Rust, Kotlin (Gradle), and PHP (Composer).
• Cross-platform tests and reusable GitHub Actions.

Quick start

Requires Python 3.11 or newer. Install the published package:

python -m pip install governed-agent-sdlc
agentkit --version
agentkit doctor

Upgrade an existing installation to this release:

python -m pip install --upgrade governed-agent-sdlc==0.10.0
agentkit --version
agentkit upgrade --dry-run
agentkit upgrade
agentkit doctor

Version 0.10.0 makes agentkit validate enforce [review].approval_mode = "github": in such a project every artifact approval must be GitHub-verified, so approvals made in local mode or recorded before 0.8.0 now fail until they are re-approved with agentkit artifact transition ... approved --approval-mode github. Projects in the default local mode are not affected. Run upgrade after upgrading the package, as since 0.9.0, and commit .agent/generated.json with the files. See release notes for the full changes.

For local development:

python -m pip install -e .
agentkit doctor
agentkit validate
python -m unittest discover -s tests -v

Initialize another project:

agentkit init ../my-project --name my-project --adapter claude-code
cd ../my-project
agentkit doctor

Use --format json with validate, doctor, generation, migration, and artifact inspection commands for automation. Preview additive initialization or adapter generation with --dry-run.

Initialization is additive: existing AGENTS.md, manifest, core policy, hook, and adapter files are not overwritten. A fresh project receives the stack profiles and a minimal AGENTS.md so generated roles always have the instructions they reference.

agentkit doctor also compares the guards under hooks/ and the files of each adapter in use with what the installed version writes, and exits 1 when one is outdated or missing, so CI can catch a project left behind by a package upgrade; a file that differs without a record only warns. agentkit upgrade rewrites those files and keeps the ones edited by hand unless --force is given; --dry-run shows what it would change. Generated files are recorded by hash in .agent/generated.json; commit it with them. See the CLI reference.

Name the stack when you initialize, and the manifest starts with its profile and its commands instead of the generic one:

agentkit init my-service --profile java
agentkit init my-app --profile flutter --profile node   # several: the first to name a command wins

The profiles are generic, python, dotnet, nuxt, node, java, flutter, go, rust, kotlin and php. The commands a profile gives are a starting point written into [commands] of the manifest: a Gradle project replaces the Maven ones of java, and flutter names no build because which artefact to build is the project's to say. kotlin runs ./gradlew check, which runs ktlint or detekt only when the build applies them, and php names composer validate and PHPUnit; a project using PHPStan or PHP CS Fixer adds those commands. The go format check is a POSIX shell line, because gofmt -l lists files without failing.

Create and move a governed artifact:

agentkit artifact new spec add-search
agentkit artifact transition docs/agent/specs/<file>.md awaiting_approval
agentkit artifact transition docs/agent/specs/<file>.md approved \
  --approved-by "github:maintainer" \
  --evidence "https://github.com/org/repo/issues/123#issuecomment-..."

An agent must never supply approval metadata for itself. The human supplies the actor and durable evidence, and CI validates that the fields exist. By default ([review].approval_mode = "local", as for review approve) the approval is recorded as given and marked verification = "unverified". With approval_mode = "github", or --approval-mode github on the command, the CLI checks through gh that the review or comment URL is in this repository and was written by a human collaborator, and records that login (github:LOGIN) as the approver. In a github project, agentkit validate rejects any approval not recorded that way; see artifact approval verification.

Project manifest

agentkit.toml or .agent/project.toml is the source of truth:

version = 1
protected_areas = ["authentication", "database-schema", "billing"]

[project]
name = "commerce"
topology = "monorepo"

[[repositories]]
id = "backend"
path = "services/api"
profiles = ["dotnet"]

[[repositories]]
id = "frontend"
path = "apps/web"
profiles = ["nuxt"]

[workflow]
require_spec = true
require_plan = true
require_review = true
require_qa = true

Repository paths are resolved from the manifest and must remain inside the project root. No machine or user-specific absolute path belongs in committed configuration.

Validation is strict and dependency-free. It rejects unsupported manifest versions and topology, missing workflow flags, invalid field types, duplicate repository identities or paths, escaping repository paths, and profile capabilities that are not provided by profiles/, adapters/, or the repository's GitHub integration.

Artifact validation also checks kind-specific parent gates, approval history for approved/active/ completed/superseded states, repository and protected-area references, timestamps, supersession, and metadata types. Artifact creation refuses filename collisions. Transitions made through the CLI are restricted to Markdown files below the active project's docs/agent/ directory.

Design boundaries

• core/ is vendor-neutral and is the only source of workflow meaning.
• profiles/ contain stack-specific commands and exclusions.
• adapters/ and generated tool files translate the core; they do not redefine it.
• hooks/ enforce deterministic safety rules. Prompts explain behavior but are not security controls.
• Hook input is fail-closed: malformed input is denied, as are credential paths/environment dumps, force pushes, and direct protected-branch refspecs.
• Hooks also hold the workflow: a role is spawned only from artifacts that passed their gate, and a role that does not own a task file may change only its completion evidence.
• docs/agent/ holds project artifacts, not hidden agent memory.

See Architecture, Workflow, and Adopting Governed Agent SDLC.

Maturity

The project is an alpha. Claude Code, Codex, OpenCode and Antigravity consume the same core contracts rather than introducing parallel policy files. The credential and git guards hold on all four; the role write scope and the role gate are enforced on Claude Code only, because only its hook input says which role is acting.

See CHANGELOG.md for release history and upgrade notes.

License

MIT. See LICENSE.