Tool-neutral core
Roles, transitions, approvals, and security policy stay independent of any vendor. Adapters carry them to Claude Code, Codex, OpenCode, and Antigravity.
Governed AI delivery
One workflow policy for Claude Code, Codex, OpenCode and Antigravity, with human approvals that can be checked against GitHub reviews and comments. A dependency-free toolkit with explicit scope, approval, review, QA, and publish gates.
$ agentkit validate Validation complete: 0 error(s), 0 warning(s). $ agentkit artifact transition docs/agent/specs/20261109162915-add-search.md approved ERROR: Approval requires --approved-by and --evidence; agents cannot self-approve. $ echo '{"tool_name":"Bash","tool_input":{"command":"git push --force origin main"}}' | python hooks/guard_git.py BLOCKED: force push is forbidden
The missing control plane
Prompts can explain good behavior. They cannot prove approval, prevent a force push, or preserve a review trail. Governed Agent SDLC turns those expectations into portable policy, structured artifacts, deterministic validation, and safety hooks.
Project policy keeps scope, credentials, merge, and release under human authority. Generated role guidance and hooks constrain supported tool actions; platform permissions remain part of the security boundary.
One accountable path
Each handoff produces readable evidence. Every gate is explicit.
Separate facts, assumptions, scope, and risk.
Build an ordered task graph from approved intent.
Work from an active task with an explicit write boundary.
Assess independently; never repair while reviewing.
Run the declared QA matrix and record evidence.
Open a PR only after a final human decision.
Policy that travels
Roles, transitions, approvals, and security policy stay independent of any vendor. Adapters carry them to Claude Code, Codex, OpenCode, and Antigravity.
Readable Markdown and TOML frontmatter make lifecycle state machine-checkable. Approvals can be verified on GitHub: only a review or comment by a human collaborator counts. agentkit review runs an OpenCodeReview or host CLI review, judges its findings against the manifest's policies, and keeps every attempt and human decision as append-only evidence.
Supported hooks deny known credential access, force pushes, protected-branch refspecs, recursive deletes outside the workspace, role writes outside their grant, and spawning a role from an artifact that has not passed its gate.
Initialize, generate, validate, diagnose, migrate the manifest, and transition artifacts with Python 3.11+. After a package upgrade, agentkit upgrade refreshes the generated guards and adapter files and keeps hand edits. In CI, the reusable validation workflow runs the toolkit revision your workflow pins.
Start with generic, Python, Node.js, Java, .NET, Nuxt, Flutter, Go, Rust, Kotlin, and PHP defaults without changing governance.
Continuously checked across Windows, Ubuntu, macOS, and three Python versions.
Trust boundaries by design
The workflow kernel owns meaning. Profiles add stack knowledge. Adapters translate policy. Hooks enforce selected high-value boundaries without replacing platform permissions.
Read the architectureThree commands to begin
Install the published package, then initialize an existing project. Nothing project-owned is overwritten.
$ python -m pip install governed-agent-sdlc
$ agentkit init ./my-project --name my-project
$ agentkit doctor ./my-project
Human authority. Agent velocity.