govkit 0.11.1

Governed AI delivery kit — spec-driven scaffolding for AI coding agents

govkit — Governed AI Delivery

AI coding agents are powerful — but without constraints, they drift. They invent architecture, skip tests, ignore NFRs, and make decisions that belong to your team. govkit puts the agent inside a governed system where your architecture contracts, acceptance criteria, and evaluation thresholds are the source of truth, not the agent's training data.

pip install govkit
govkit apply --agent claude-code --target .
govkit calibrate

Install govkit, apply it to your project, then calibrate the defaults to match your repo. From there the governance workflow your team follows is what keeps the agent aligned — every feature, every time.

govkit works with any project language — Python, C#, Java, Go, TypeScript, or anything else. It copies Markdown specs, YAML configs, and Gherkin feature files into your project directory. Python is a dev-machine tool requirement only; it is not added as a project dependency.

---

Get started in 4 steps

You need Python 3.11+ on your machine (govkit is a dev tool, never a project dependency) and an AI coding agent — Claude Code, GitHub Copilot, or OpenAI Codex. govkit ships the governance configuration; you bring the agent.

Work top to bottom. Steps 1–2 install governance; steps 3–4 make it govern your project, not a generic one.

1. Install govkit

pip install govkit

2. Apply it to your project

From your project root:

govkit apply --agent claude-code --target .

govkit detects your stack (language, framework, CI), scaffolds the agent rules + architecture contracts, and writes a .govkit/ marker recording your configuration. Prefer to set everything explicitly, or use a different agent/type/level? See Choose your options.

3. Calibrate the defaults to your repo

[!IMPORTANT] This step is not optional. The files govkit installs are sound generic defaults — and your agent treats them as law. If you skip calibration, the agent governs your project against someone else's architecture decisions.

govkit calibrate

govkit calibrate walks you through a 9-step review — installed configuration, tech stack, architecture boundaries, API/query conventions, testing + BDD policy, agent guidance, rules, CI gates, and skill context. For each step it shows the installed default next to what it detected in your repo, then you confirm or modify. Prefer to review with the team first? govkit calibrate --non-interactive writes a GOVKIT_CALIBRATION_CHECKLIST.md todo file instead.

💡 Let the agent help — but you decide. You can ask the agent to draft edits — "update TECH_STACK.md to match our pyproject.toml" or "align BOUNDARIES.md with our actual folder layout" — pointing it at evidence you already have (dependency manifests, existing code). You confirm every change. These files are the source of truth that governs the agent; don't let it invent its own guardrails from its training data — that's the exact drift govkit exists to prevent.

4. Commit your governance baseline

git add .govkit CLAUDE.md docs/ governance/ ci/
git commit -m "chore: add govkit governance baseline"

You're ready to build. See The feature lifecycle for how you drive the agent through a feature, and run govkit doctor once you have source code (and in CI) to catch governance that has drifted out of sync with your repo.

Upgrading later? pip install --upgrade govkit && govkit upgrade --target . refreshes the files govkit owns without touching the contracts you've customized. See Keeping contracts up to date.

---

Choose your options

Each govkit apply configures one project shape. Pick one value per flag:

Flag	Options	Pick this if…
--agent	claude-code · copilot · codex	…matches the AI tool your team uses
--type	api · cli · ui-react · ui-angular · data	…describes this repo (or subdir) — backend service, CLI, React/Angular UI, or dbt data project
--level	3 · 4 · 5	…3 governed foundations (default) · 4 spec-driven delivery · 5 GenAI operations — see Maturity Levels
--ci	github · azure	…your CI platform
--stack	python-fastapi · dotnet-aspnet · java-spring-boot · nodejs-fastify · go-gin · python-dbt	…backend/data only; auto-detected, defaults to python-fastapi. See Switching Tech Stacks

Running govkit apply with no --type/--ci/--stack flags prompts interactively and auto-detects sensible defaults from your repo. A .govkit marker records every choice so later commands (calibrate, validate, upgrade, doctor) need no re-specification.

Full example commands for every combination

# Level 3: Governed AI Delivery (Foundations) — agent rules + architecture docs only (default)
govkit apply --agent claude-code --level 3 --type api --ci github --target .

# Level 4: Spec-Driven Add-On — adds the features/ directory and 5-artifact contract
govkit apply --agent claude-code --level 4 --type api --ci github --target .

# Level 5: GenAI Operations (LLM routing, evaluation, guardrails)
govkit apply --agent claude-code --level 5 --type api --ci github --target .

# Python CLI tool + Azure DevOps
govkit apply --agent copilot --type cli --ci azure --target .

# React UI project (L4)
govkit apply --agent copilot --level 4 --type ui-react --ci github --target .

# Angular UI project (L4) on OpenAI Codex
govkit apply --agent codex --level 4 --type ui-angular --ci github --target .

# Data project (L4) — dbt-layered, python-dbt stack inferred from dbt_project.yml
govkit apply --agent claude-code --level 4 --type data --ci github --target .

# Pick a non-default backend stack
govkit apply --agent claude-code --level 4 --type api --ci github --stack dotnet-aspnet --target .

Fullstack-in-one-repo is supported via the monorepo pattern — one govkit apply per subdirectory. There is no --type fullstack.

What gets installed

After applying, your project contains artifacts appropriate to the shape you picked.

Backend shape (--type api or --type cli):

your-project/
├── CLAUDE.md (or .github/copilot-instructions.md, or AGENTS.md)
├── .claude/rules/ (or .github/instructions/, or nested AGENTS.md per layer)
│   └── api.md, services.md, ports.md, adapters.md, security.md, repo-scope.md
├── .claude/skills/ (or .github/skills/, or .agents/skills/)
│   └── architecture-preflight/, spec-planning/, implementation-plan/, adr-author/
├── docs/backend/
│   ├── architecture/   — ARCH_CONTRACT, API_CONVENTIONS, TECH_STACK, etc.
│   └── evaluation/     — eval_criteria.md, scoring rubrics
├── features/           — created empty at L4+; scaffold your first feature with govkit init
├── governance/backend/schemas/   — eval_criteria.schema.json
└── ci/github/ (or azure/)        — l3-quality-gate.yml + L4/L5 gates

UI shape (--type ui-react or --type ui-angular):

your-project/
├── CLAUDE.md (or .github/copilot-instructions.md, or AGENTS.md)
├── src/CLAUDE.md       — Claude only: nested UI layer rules under src/
│                         (Codex uses src/AGENTS.md + nested AGENTS.md per layer;
│                          Copilot uses .github/instructions/ with src-scoped applyTo: globs)
├── .claude/rules/ (or .github/instructions/)
│   └── repo-scope.md, test-first.md (L4+), spec-compliance.md (L4+)
├── .claude/skills/
│   └── ui-architecture-preflight/, ui-spec-planning/, ui-implementation-plan/, ui-adr-author/
├── docs/ui/
│   ├── architecture/   — MVVM_CONTRACT, ACCESSIBILITY_STANDARDS, react|angular subdirs
│   └── evaluation/     — eval_criteria.md, scoring rubrics
├── governance/ui/      — schemas, templates
└── ci/github/ (or azure/)  — l3-ui-quality-gate.yml + L4/L5 UI gates

Backend installs ship no UI artifacts; UI installs ship no backend artifacts. The CI dispatch is type-aware: backend types get l3-quality-gate.yml, UI types get l3-ui-quality-gate.yml. For fullstack monorepos, run one govkit apply per app subdirectory — see the monorepo pattern.

Starter templates and worked examples are bundled inside the govkit package, not copied into your project by govkit apply. Use govkit init <feature-name> to scaffold a new feature from the appropriate starter, or run govkit list to see what is available.

---

Commands

Command	What it does
govkit apply	Install / scaffold governance into your project. Detects your stack, writes the .govkit marker.
govkit calibrate	Guided 9-step review to make the installed generic defaults match your repo. --non-interactive writes a checklist file; --only <step> revisits one decision.
govkit doctor	Read-only governance-fit checks (rule globs resolve, CI/stack/language match, stale baselines, extension manifests). Run once you have source code, and in CI. Monorepo-aware.
govkit validate	Level-aware per-feature compliance check (artifact existence, Gherkin structure, NFR coverage, eval-criteria schema, prediction thresholds). No-op at L3.
govkit init <feature>	Scaffold a new feature folder from the appropriate starter (L4+).
govkit stack	stack list shows bundled tech-stack overlays; stack apply <id> swaps the stack on an existing install.
govkit upgrade	Refresh the files govkit owns (contracts, CI gates, templates) to a new version without touching the files you own.
govkit list	List available agents and starter templates.

govkit doctor and govkit validate cover different surfaces: doctor checks whether the installed governance still fits the repo; validate checks whether your features meet the governed contract. Both are designed to run in CI.

---

Maturity Levels

govkit supports three operating levels in an additive ladder. Each level commits the team to a different way of working, not just a different artifact count. The boundary between L3 and L4 sits at the most meaningful place: whether the team adopts a features/ directory model and per-feature spec contracts.

Level	Name	What it ships	What the team commits to
Level 3	Governed AI Delivery (Foundations)	Agent rules, architecture contracts under docs/*/architecture/, /adr-author skill, lean CI gate (commit-format + import-linter + sonar/snyk). No features/ directory, no per-feature artifacts.	"Our AI agents follow our architecture contracts." Lowest-friction entry; no project-structure change required.
Level 4	Spec-Driven Add-On	Adds the features/<name>/ 5-artifact contract (acceptance.feature, nfrs.md, eval_criteria.yaml, plan.md, architecture_preflight.md), feature-coupled skills (/spec-planning, /architecture-preflight, /implementation-plan), test-first + spec-compliance rules (binding), governance CI jobs, and per-feature evaluation prediction (FIRST + 7 Virtues, average ≥ 4.0).	"We adopt spec-first, test-first feature delivery on top of L3." govkit init becomes meaningful here.
Level 5	GenAI Operations	LLM-specific NFR categories (latency, cost, fallback, safety), agent_topology.md for multi-agent features, deepeval/promptfoo/guardrails CI gates, LLM gateway/observability/multi-agent rules, LiteLLM routing, OpenLLMetry + Langfuse, RAGAS, NeMo Guardrails.	"Our LLM features are governed (routing, evaluation, safety)." Builds on L4.

Start at Level 3 (default) if you want governed AI agents without restructuring your codebase. Move to Level 4 when your team is ready to commit to spec-first feature delivery. Move to Level 5 when shipping LLM-powered features that need governed model routing, evaluation, and safety.

The ladder is additive: L4 ⊃ L3, L5 ⊃ L4. Files installed at lower levels are not replaced by higher levels (with one exception: the agent's top-level entry point — CLAUDE.md / .github/copilot-instructions.md / AGENTS.md — is re-issued at each level so the agent sees the right operating mode).

Level 3 — Governed AI Delivery (Foundations)

Your AI agent operates aligned to your architecture contracts:

• Reads docs/<area>/architecture/ (ARCH_CONTRACT, BOUNDARIES, TESTING, SECURITY_AUTH_PATTERNS, etc.) on every turn
• Path-scoped rules trigger when editing files in each layer (api/, services/, ports/, adapters/, security/)
• ADRs required for any standard extension, override, or boundary change — scaffolded with /adr-author
• Test-first is recommended (the binding rule lives at L4)
• CI quality-gate enforces commit-format, import-linter (architecture boundaries), SonarQube, Snyk
• No features/ directory is created. govkit init errors at L3 with a pointer to --level 4.
• govkit validate is a no-op (returns 0 with informational message).

Level 4 — Spec-Driven Add-On

Everything in Level 3, plus:

Every feature lives under features/<name>/ with five required artifacts:

• Defined with Gherkin acceptance criteria tagged to NFR categories
• Constrained with fully populated NFRs (no TBD entries permitted)
• Governed by evaluation criteria validated against a JSON Schema
• Planned through Architecture Preflight + Spec Planning + Implementation Plan skills
• Bounded by hexagonal architecture contracts with FIRST + 7 Virtues prediction (average ≥ 4.0)
• Enforced by governance CI jobs (artifact existence, eval-criteria schema, prediction thresholds, contract compatibility)

Level 5 — GenAI Operations

Everything in Level 4, plus:

• Routed through LiteLLM as the sole LLM gateway (model routing, fallback, cost tracking)
• Observed via OpenLLMetry + Langfuse (LLM-specific telemetry, trace storage, prompt versioning)
• Evaluated with DeepEval (quality metrics), Promptfoo (adversarial testing), and RAGAS (retrieval evaluation)
• Guarded by NeMo Guardrails (conversational safety) and Guardrails AI (structured output validation)
• Extended with LLM-specific NFRs (latency, cost, fallback, safety) and 3 additional CI gates

The AI agent operates inside a governed system. Architecture, evaluation, and feature artifacts are the source of truth — not the agent.

Migrating from v0.6.x

The L3 / L4 maturity-model meaning changed in v0.7.0. If your .govkit marker says level: "3" (the v0.6 simpler 3-artifact model), run:

govkit upgrade --migrate-levels --target .

You'll be prompted to choose: migrate your existing 3-artifact features to the new L4 (with stub generation for the two new artifacts), or adopt new-L3 (no features/). See CHANGELOG.md for the full migration guide.

---

The feature lifecycle

Once govkit is installed and calibrated, here is how you interact with the agent to deliver a feature. This lifecycle applies to every feature, regardless of project type. The commands below use Claude Code; see Agent command equivalents for Copilot and Codex.

Step 1: Create a feature folder

govkit init my_feature --target .

Or specify the starter type explicitly:

govkit init my_feature --starter backend --target .

The command auto-detects your maturity level from .govkit. govkit init requires Level 4 or higher — Level 3 (Foundations) ships agent rules and architecture contracts only and has no features/ directory model. Running govkit init at L3 errors with a pointer to govkit apply --level 4. At L4 the bundled starter has all 5 artifacts; at L5 the starter adds agent_topology.md for multi-agent features.

For Level 4 projects, each starter's eval_criteria.yaml includes mode selection instructions at the top. Set the mode field to match your feature type: llm (LLM generation/retrieval), deterministic (pure logic), or none (configuration artifacts). If the mode is deterministic or none, delete the llm_evaluation section.

Step 2: Write your acceptance criteria

Edit features/my_feature/acceptance.feature with your Gherkin scenarios:

• Write happy path and failure/edge case scenarios
• Tag NFR scenarios with @nfr-performance, @nfr-security, etc.
• Tag E2E scenarios with @e2e (UI projects)
• Add @contract scenarios if the feature produces shared artifacts

Step 3: Complete your NFRs

Edit features/my_feature/nfrs.md — replace every TBD entry with concrete requirements. The agent will refuse to proceed if any TBD entries remain.

Step 4: Run architecture preflight

Ask the agent to validate your feature against the architecture contracts with /architecture-preflight my_feature. The agent produces architecture_preflight.md covering boundary analysis, security impact, evaluation impact, and whether an ADR is needed. If an ADR is required, create it next with /adr-author my_feature.

Step 5: Generate the plan

Ask the agent to create the implementation plan with /spec-planning my_feature. The agent generates plan.md and eval_criteria.yaml. The plan includes:

• Increments with deliverables and tests
• An Evaluation Compliance Summary predicting FIRST and 7 Virtue scores
• Each increment sized as a single committable unit (~300 lines)

The agent will not proceed if predicted averages are below 4.0.

Step 6: Review the implementation plan

Ask the agent to break the plan into a detailed task checklist with /implementation-plan my_feature. Review and approve before implementation begins.

Step 7: Implement incrementally

Work through the plan one increment at a time. For each increment:

1. The agent writes production code respecting architecture boundaries
2. The agent writes tests (unit, integration, contract as needed)
3. You review and commit: feat(my_feature): increment 1 — <name>
4. Move to the next increment

Do not skip increments or combine multiple increments into one commit.

Step 8: Push and merge

Open a PR. CI gates automatically run:

• Schema validation of eval_criteria.yaml
• FIRST and 7 Virtue prediction completeness
• Unit, component, and E2E tests
• Architecture boundary enforcement
• Security scan and quality gates
• Accessibility checks (UI projects)

All gates must pass before merge.

Agent command equivalents

The lifecycle is identical across agents; only the invocation syntax differs.

Step	Claude Code	Copilot	Codex
Architecture preflight	/architecture-preflight my_feature	/architecture-preflight	$architecture-preflight my_feature
Author ADR	/adr-author my_feature	/adr-author	$adr-author my_feature
Spec planning	/spec-planning my_feature	/spec-planning	$spec-planning my_feature
Implementation plan	/implementation-plan my_feature	/implementation-plan	$implementation-plan my_feature

Copilot infers the feature from context rather than taking it as an argument; Codex invokes skills with a $ prefix.

---

Project Type Details

The 8-step lifecycle above applies to all project types. Key differences by type:

Backend API

Architecture: Hexagonal Architecture — ports and adapters. API routes are the inbound adapter layer. See docs/backend/architecture/API_CONVENTIONS.md.

Layer rules (load automatically when editing files):

• api.md for **/api/**
• services.md for **/services/**
• ports.md for **/ports/**
• adapters.md for **/adapters/**
• security.md for **/security/** and **/auth/**

CI gates: ci/github/quality-gate.yml, ci/github/eval-gate.yml (or ci/azure/ for Azure DevOps)

CLI

Architecture: Hexagonal Architecture — CLI commands are the inbound adapter layer (same position as API routes). See docs/backend/architecture/CLI_CONVENTIONS.md.

Layer rules (load automatically):

• cli.md for **/cli/** and **/commands/**
• services.md, ports.md, adapters.md, security.md as above

CI gates: Same backend gates — ci/github/quality-gate.yml, ci/github/eval-gate.yml

React UI

Architecture: MVVM with vertical slice feature structure. Layer order is API → ViewModel → View. See docs/ui/architecture/MVVM_CONTRACT.md.

Layer rules (load automatically):

• components.md for View layer
• viewmodel.md for hooks and store
• api.md for API client functions
• accessibility.md for accessibility concerns

Implementation order: API functions → React Query hooks → Zustand stores → Components → E2E tests

CI gates: ci/github/ui-quality-gate.yml, ci/github/ui-eval-gate.yml

Angular UI

Architecture: MVVM with vertical slice feature structure. Standalone components with OnPush. See docs/ui/architecture/MVVM_CONTRACT.md.

Layer rules: Same as React, with Angular-specific content.

Implementation order: API functions → TanStack Query inject functions → Signal stores → Standalone components → E2E tests

CI gates: ci/github/ui-quality-gate.yml, ci/github/ui-eval-gate.yml

Data (dbt)

Architecture: dbt-layered — Staging → Intermediate → Marts. Staging cleans source data (one source per model), Intermediate holds joins and business logic, Marts are the downstream contracts. See docs/data/architecture/BOUNDARIES.md and the python-dbt stack overlay's MODEL_LAYERING.md.

Layer rules (load automatically):

• staging.md for **/models/staging/**
• intermediate.md for **/models/intermediate/**
• marts.md for **/models/marts/**
• data-quality.md for **/tests/** (dbt schema + singular tests)
• pii.md for **/models/**, **/macros/**, **/seeds/** (PII tagging + masking)

Stack overlay: python-dbt (dbt-core + Snowflake / BigQuery / Redshift / Postgres adapter, SQLfluff, dbt schema tests, optional dbt-expectations for L4).

Worked starter: govkit init <feature> --starter data scaffolds a customer_dim_freshness feature with @nfr-freshness / @nfr-quality / @nfr-pii / @nfr-lineage / @nfr-reliability / @nfr-cost Gherkin scenarios.

CI gates: intentionally empty in this release — gate selection (source freshness, dbt test, SQLfluff) is the kind of thing data teams want to shape themselves. Wire your own in ci/github/ or ci/azure/ for now; opinionated defaults arrive in a future release.

Agent support: claude-code only for now. copilot and codex variants follow once the shape is validated against real dbt teams.

---

Monorepo (Fullstack) Pattern

Each govkit apply configures one project shape. There is no --type fullstack. Teams that need both backend and UI in a single repo run govkit apply once per subdirectory:

govkit apply --agent claude-code --type api      --level 4 --ci github --target apps/api
govkit apply --agent claude-code --type ui-react --level 4 --ci github --target apps/web

Each subdir becomes a self-contained govkit install — separate .govkit marker, separate features/, separate CI gates. The three agents all support subpath governance natively:

• Claude Code — recursive CLAUDE.md discovery picks up the right shape based on the open file's directory
• Codex — directory-walk loader concatenates AGENTS.md from leaf to root
• Copilot — applyTo: globs in each instructions file (one small post-install adjustment to prefix the app path so globs don't cross app boundaries)

govkit calibrate and govkit doctor are monorepo-aware: run them with no --target from the repo root and they discover every .govkit/ install under the tree and process each app in turn.

For the complete setup — directory layout, CI workflow examples, the Copilot applyTo: prefix tweak, feature governance per app, upgrade flow, and gotchas — see docs/MONOREPO_PATTERN.md.

If your backend and UI live in separate repositories instead of subdirectories of a monorepo, see Multi-Repository Features below — different coordination problem.

---

Switching Tech Stacks

GovKit ships stack overlays — small bundles of 6 stack-specific architecture docs plus metadata. Pick one at install time with --stack, or swap later with govkit stack apply. Stack overlays apply to backend types (api, cli) and the data type; UI installs ignore --stack. The 6 docs vary per shape:

• Backend stacks (python-fastapi, dotnet-aspnet, java-spring-boot, nodejs-fastify, go-gin): TECH_STACK.md, API_CONVENTIONS.md, TESTING.md, LAYER_IMPLEMENTATION.md, SECURITY_AUTH_PATTERNS.md, OBSERVABILITY_PORT_CONTRACT.md.
• Data stacks (python-dbt): TECH_STACK.md, QUERY_CONVENTIONS.md, TESTING.md, MODEL_LAYERING.md, PII_HANDLING.md, LINEAGE_OBSERVABILITY.md.

See what's available

govkit stack list

Lists every bundled overlay (id, display name, summary) along with the apply commands.

Pick a stack at install time

govkit apply --agent claude-code --target . --level 4 --type api --ci github \
             --stack dotnet-aspnet

If you omit --stack, govkit detects your stack from the repo and falls back to the python-fastapi overlay, recording the choice as an assumption in .govkit/marker.json so govkit doctor can warn if it doesn't fit your repo.

Swap stacks on an existing install

govkit stack apply java-spring-boot --target .

Re-applies the new overlay on top of the existing install. Edit-protection respects user changes: any of the 6 stack docs you've modified since the last apply are preserved unless you pass --force (your edits are detected via the govkit:editable header + file mtime vs. the marker's applied_at).

Why only 6 files?

The agent rules and most architecture docs (DESIGN_PRINCIPLES, ARCH_CONTRACT, BOUNDARIES, GHERKIN_CONVENTIONS, ERROR_MAPPING, etc.) are language-agnostic and ship from the baseline. Only these 6 vary per stack:

File	What it defines
TECH_STACK.md	Languages, versions, approved frameworks
API_CONVENTIONS.md	Route patterns and request/response idioms
TESTING.md	Test framework, mocking library, BDD tool
LAYER_IMPLEMENTATION.md	DI patterns, interface idioms, DTO style
SECURITY_AUTH_PATTERNS.md	Auth libraries, token handling, hashing
OBSERVABILITY_PORT_CONTRACT.md	Structured logging library, OTel SDK

Bundled stacks

Id	Shape	Stack
python-fastapi	backend	Python 3.11+ / FastAPI / pytest (default for api / cli)
dotnet-aspnet	backend	C# 12 / .NET 8 / ASP.NET Core Minimal APIs / xUnit
java-spring-boot	backend	Java 21 / Spring Boot 3 / Spring Web MVC / JUnit 5
nodejs-fastify	backend	Node.js 20 LTS / TypeScript 5 / Fastify 4 / Vitest
go-gin	backend	Go 1.22+ / Gin / standard library testing + testify
python-dbt	data	Python 3.11+ / dbt-core (staging → intermediate → marts) / Snowflake | BigQuery | Redshift | Postgres adapter / SQLfluff / dbt schema tests (default for data)

After applying a stack, review the installed files and adapt anything specific to your repo (approved library versions, internal service names, etc.) — govkit calibrate walks you through this. GOVKIT_SETUP_REVIEW.md at the target root lists each stack doc with a one-line review prompt. Consider raising an ADR to document the stack decision.

See cli/stacks/README.md for the complete guide, including how to add new stacks.

---

Extensions

Govkit supports optional extension packs that layer additional architecture contracts and governance templates on top of the core kit. They're a drop-in mechanism: there is no govkit extension add command — the folder under extensions/<id>/ in your project is the install.

How to add an extension

1. Create the folder. Extensions live at the root-level extensions/ directory of your project — a sibling of docs/, governance/, and features/:

   <project>/
   ├── docs/
   ├── governance/
   ├── features/
   ├── extensions/
   │   └── <extension-id>/
   │       ├── manifest.yaml
   │       ├── README.md
   │       ├── docs/
   │       └── governance/
   └── .govkit
   

2. Drop in the extension's files. Clone the extension repo, copy its folder, or vendor it into extensions/<id>/. Everything the extension ships (its manifest.yaml, docs/, and governance/) lives in-place under that folder — all manifest paths are relative to it.

3. Re-run govkit apply. It scans extensions/*/manifest.yaml and prints each extension it discovers, so you get confirmation it was picked up.

4. Validate it. govkit validate and govkit doctor (checks D013/D014) verify the manifest and flag any undeclared overlap with core contracts (see below).

When extensions/ is absent, govkit behaves exactly as it does without extensions — they are entirely optional.

See extensions/agentic-skills/ in this repository for a complete reference example.

Authoring an extension

Each extension is self-describing: it declares its own id, version, contract_sets, capabilities, and agent guidance in its manifest. Govkit needs no per-extension code.

Minimal manifest (extensions/<id>/manifest.yaml):

id: my-extension
name: My Extension
version: 0.1.0
extension_type: architecture
contract_sets:
  - id: my_contracts
    description: ...
    paths:
      - docs/backend/architecture/MY_CONTRACT.md
    capabilities:
      - my-capability

The extension id must match its folder name and the pattern ^[a-z0-9][a-z0-9-]*$. Every path listed in contract_sets[].paths (and templates[].path) must exist as a file under the extension folder — govkit validate reports missing or out-of-bounds paths.

Resolving overlap with core contracts

When an extension contract covers the same topic as a core govkit contract (e.g. an AGENT_EVALUATION_CONTRACT.md extension alongside core EVALUATION_LLM_CONTRACT.md), the manifest declares the relationship explicitly via relates_to:

contract_sets:
  - id: my_contracts
    paths: [docs/backend/architecture/AGENT_EVALUATION_CONTRACT.md]
    relates_to:
      extends:    [docs/backend/architecture/EVALUATION_LLM_CONTRACT.md]   # both apply; stricter rule wins
      supersedes: []                                                        # extension replaces core (requires ADR)

• relates_to.extends — the extension layers additional constraints on top of the core contract. The agent reads both and applies whichever is stricter on any specific point.
• relates_to.supersedes — the extension replaces the listed core contract for rules in its scope. Requires an ADR in the consuming project.

Undeclared overlap is detected. govkit validate and govkit doctor run a filename-token heuristic: if an extension contract shares a topic token with a core contract under docs/backend/architecture/ and relates_to does not declare the relationship, the validator emits a WARN (or FAIL under --strict) asking the extension author to declare the intent. This prevents silent drift when extension authors and core authors update the same topic area independently.

Agent reading order at preflight time. The architecture-preflight skill reads core contracts first, then extension contracts; it prefers extensions only when supersedes is declared. If an applicable extension and a core contract conflict and relates_to does not declare the relationship, the agent halts and requests either a manifest update or an ADR rather than silently picking one.

---

Multi-Repository Features

If your feature spans multiple repositories (e.g., Auth Service + Client SDK + API Gateway), see:

• CROSS_REPO_FEATURES.md — Complete guide to planning, implementing, and testing features across repos
• REPO_SCOPE_ANALYSIS_GUIDANCE.md — How to declare repo ownership in your feature spec
• features/example-jwt-unification/ — Worked example of a 3-repo JWT authentication feature

The key principle: Every feature must declare which repositories own which parts. This prevents agents from writing code in the wrong repo.

Multi-Repo FAQ

Q: My feature needs changes in Auth Service, API Gateway, and Frontend. Where do I document this? A: In the primary owner repo's features/<feature>/nfrs.md, add a "Repository Scope" section with a table listing each repo, owner team, modules, and contracts. See CROSS_REPO_FEATURES.md#repository-ownership-table.

Q: Can we implement the feature in just one repo and copy code to the others later? A: No — this violates ownership and creates duplication. Each repo implements its own portion against the shared contract. See CROSS_REPO_FEATURES.md#common-pitfalls.

Q: Should we wait for Repo A to finish before Repo B starts? A: No. Each repo implements in parallel using mocks for external dependencies. Only the final integration tests (after all repos merge) verify cross-repo contracts. See CROSS_REPO_FEATURES.md#implementation-stage-parallel.

Q: How do we test a feature that depends on another repo's code? A: Each repo has unit tests (mocking externals) and contract tests (verifying its own implementation). After all repos merge to main, run integration tests to verify end-to-end behavior. See CROSS_REPO_FEATURES.md#testing-strategy.

Q: What if the repos have deployment dependencies (one must be live before the other)? A: Document the order in your nfrs.md "Key Cross-Repo Contracts" section. Ideally, design contracts to be backward-compatible so deployment order is flexible. See CROSS_REPO_FEATURES.md#integration-stage-sequential.

---

Keeping contracts up to date

When you upgrade govkit, run govkit upgrade to refresh the files that govkit owns — architecture contracts, CI gate pipelines, plan templates — without touching the files your team owns.

pip install --upgrade govkit
govkit upgrade --target .

govkit distinguishes three categories of files:

Category	Examples	apply	upgrade
Agent config	CLAUDE.md, .claude/rules/, .agents/skills/	Always overwrite	Always overwrite
Governed contracts	docs/backend/architecture/, governance/backend/templates/, ci/github/	Write once (skip if present)	Overwrite
Project artifacts	features/starter_*/, your ADRs, filled-in feature files	Write once (skip if present)	Skip

After upgrading, review the diff and commit:

git diff
git add -p
git commit -m "chore: upgrade govkit governance contracts to vX.Y.Z"

Use --force to re-apply even when the version is already current — useful for resetting a contract file to govkit defaults after an accidental edit:

govkit upgrade --target . --force

---

Interpreting Validation Failures

When govkit validate --target . reports failures, here's what they mean and how to fix them:

Failure	Meaning	Fix
acceptance.feature not found	Feature folder is missing its Gherkin spec	Copy from starter and write scenarios
no Feature: keyword or no Scenario: keyword	Gherkin file exists but is malformed	Add Feature: header and at least one Scenario: with Given/When/Then
nfrs.md contains TBD entries	NFR categories have placeholder values	Replace every TBD with a concrete, measurable requirement
eval_criteria.yaml missing or invalid	Eval config doesn't match the JSON Schema	Check governance/*/schemas/eval_criteria.schema.json for required fields
plan.md missing evaluation_prediction	Plan exists but has no prediction block	Add the evaluation_prediction YAML block (see worked examples)
predicted FIRST average below 4.0	Predicted test quality is below threshold	Revise the plan — improve test strategy or split complex increments
predicted Virtue average below 4.0	Predicted code quality is below threshold	Revise the plan — simplify design, reduce complexity, improve separation
NFR tag coverage incomplete	Some NFR categories lack corresponding Gherkin tags	Add @nfr-<category> tags to relevant scenarios in acceptance.feature

For governance that has drifted out of sync with your repo (rule globs matching nothing, CI/stack mismatches, stale baselines), run govkit doctor — it covers a different surface than validate.

---

Troubleshooting & FAQ

Q: govkit: command not found after installation A: Ensure your Python scripts directory is on your PATH. Try python -m cli.govkit as a fallback, or reinstall with pip install --user govkit.

Q: govkit apply fails with "no agent found" A: Check that you're using a valid agent name (claude-code, copilot, or codex). Run govkit list to see available agents.

Q: The agent ignores my architecture rules A: Verify the rules files were copied to the correct location (.claude/rules/, .github/instructions/, or the nested AGENTS.md files for Codex). Run govkit doctor — its D001 check flags any rule whose path globs resolve to zero files in your repo, which is the most common cause. Claude Code loads rules based on the file path you're editing; Codex walks the directory tree from repo root down to the current working directory and concatenates each AGENTS.md it finds.

Q: How do I update to a newer version of govkit? A: Run pip install --upgrade govkit, then govkit upgrade --target . to refresh govkit-owned files without touching your customized contracts. See Keeping contracts up to date.

Q: Can I use govkit on an existing project with existing code? A: Yes. govkit apply copies governance artifacts into your project without modifying existing code. Then run govkit calibrate to align ARCH_CONTRACT.md and the other contracts with your existing architecture rather than the defaults.

Q: What if my architecture doesn't match the Hexagonal defaults? A: Customize the architecture docs after install (calibrate's "Architecture boundaries" step walks you through it). The agent follows whatever BOUNDARIES.md / ARCH_CONTRACT.md say — if your project uses clean or layered architecture, document it there. Consider creating an ADR explaining the architectural choice.

Q: Can I use multiple agents in the same project? A: Yes. Run govkit apply once for each agent. They install to different paths (.claude/, .github/, and AGENTS.md + .agents/ for Codex) and share the same docs/, governance/, and features/ artifacts. All agents read the same architecture contracts.

Q: How do I add a new NFR category? A: Add the category as a ## Heading in your feature's nfrs.md, add corresponding @nfr-<category> tags to acceptance scenarios, and update cli/validate.py's category_to_tag mapping if you want automated tag coverage validation.

Q: The CI pipeline fails because SonarQube/Snyk isn't configured A: These tools are optional. If your team doesn't use them, remove or comment out those jobs from the CI pipeline files. See ci/README.md for details on required secrets.

Q: What does "thresholds_met: false" mean in my plan? A: Your predicted FIRST or Virtue average is below 4.0, or a predicted accessibility violation count is above zero. Revise the plan — simplify the design, improve test strategy, or split large increments before proceeding.

---

Architecture Reference

The files linked below are installed into your project by govkit apply. Links point to their source in the govkit repository for reference.

Backend (Core — Level 4)

• ARCH_CONTRACT.md — Hexagonal Architecture contract
• BOUNDARIES.md — Layer dependency rules
• API_CONVENTIONS.md — FastAPI conventions
• CLI_CONVENTIONS.md — Click/Typer conventions
• DESIGN_PRINCIPLES.md — SOLID, DRY, YAGNI, KISS
• GHERKIN_CONVENTIONS.md — NFR tags, coverage rules
• GHERKIN_TAGS.md — Standard tag reference
• SECURITY_AUTH_PATTERNS.md
• TECH_STACK.md
• TESTING.md
• AGENT_ARCHITECTURE.md — AI agent design patterns (LangGraph, tools, evaluation)
• ERROR_MAPPING.md — Domain exception to HTTP status mapping
• OBSERVABILITY_PORT_CONTRACT.md — Observability port interface
• CROSS_CUTTING_CONCERNS.md — DTOs, validation, pagination, timestamps
• ADR/TEMPLATE.md

Backend (GenAI Contracts — Level 5)

• LLM_GATEWAY_CONTRACT.md — LiteLLM as sole LLM gateway, provider abstraction, fallback, cost tracking
• OBSERVABILITY_LLM_CONTRACT.md — OpenLLMetry (telemetry emission) + Langfuse (trace storage, prompt versioning)
• GUARDRAILS_CONTRACT.md — NeMo Guardrails (conversational safety) + Guardrails AI (structured output validation)
• EVALUATION_LLM_CONTRACT.md — DeepEval (quality), Promptfoo (adversarial), RAGAS (retrieval)

Backend (Practical Guides — Level 5)

• litellm-setup.md — LiteLLM proxy config, model aliases, fallback chains
• openllmetry-setup.md — Auto-instrumentation, export to Langfuse
• langfuse-integration.md — Trace viewing, prompt management, dashboards
• deepeval-usage.md — Writing DeepEval test cases, metrics, CI integration
• promptfoo-usage.md — Adversarial configs, red-team suites
• nemo-guardrails-setup.md — Colang dialog definitions, rail configs
• guardrails-ai-setup.md — Guard definitions, validator config
• ragas-evaluation.md — RAG-specific metrics, dataset preparation

UI (Shared)

• MVVM_CONTRACT.md — framework-agnostic MVVM contract
• ADR/TEMPLATE.md

React UI

• COMPONENT_CONVENTIONS.md
• STATE_MANAGEMENT.md — React Query + Zustand
• TECH_STACK.md

Angular UI

• COMPONENT_CONVENTIONS.md
• STATE_MANAGEMENT.md — TanStack Angular Query + Signals
• TECH_STACK.md

Data (Core — Level 4)

• ARCH_CONTRACT.md — dbt-layered architecture contract
• BOUNDARIES.md — Staging / Intermediate / Marts layer rules
• DESIGN_PRINCIPLES.md — Data engineering principles (idempotency, freshness, etc.)
• PIPELINE_CONTRACT.md — Source freshness, run cadence, SLAs
• DATA_QUALITY_CONTRACT.md — Schema tests, distribution tests, severity policy
• LINEAGE_CONTRACT.md — Column-lineage requirements (especially for PII)
• PII_HANDLING_CONTRACT.md — PII tagging, masking macro contract
• ENVIRONMENTS.md — dev / ci / staging / prod separation + PII masking by env

Data (Stack overlay — python-dbt)

• TECH_STACK.md — dbt-core + warehouse adapter, SQLfluff, dbt-expectations
• QUERY_CONVENTIONS.md — CTE skeleton, stg_<source>__<table> naming
• TESTING.md — dbt schema tests, custom singular tests, freshness
• MODEL_LAYERING.md — staging / intermediate / marts materialization defaults
• PII_HANDLING.md — mask_pii() macro pattern
• LINEAGE_OBSERVABILITY.md — Exposure docs, OpenLineage hooks

---

Evaluation Reference

Standards (All Levels)

• eval_criteria.md — FIRST, 7 Code Virtues, LLM eval, scoring model
• FIRST_SCORING_RUBRIC.md — 1-5 scoring definitions per FIRST principle
• VIRTUE_SCORING_RUBRIC.md — 1-5 scoring definitions per virtue
• EVAL_STACK.md — Evaluation tooling by environment (Langfuse, DeepEval, Promptfoo, RAGAS, home-grown)

Schemas

• eval_criteria.schema.json — Validates feature eval_criteria.yaml (includes L5 eval_class values)
• evaluation_prediction.schema.json — Schema for plan.md prediction blocks (includes optional L5 llm_evaluation section)
• guardrails_config.schema.json — Validates guardrails configuration (L5)

UI Evaluation

• eval_criteria.md — FIRST, accessibility, E2E coverage
• FIRST_SCORING_RUBRIC.md — UI-adapted FIRST scoring
• VIRTUE_SCORING_RUBRIC.md — UI-adapted virtue scoring
• eval_criteria.schema.json

Evaluation by Level

Level	What's Evaluated	Tools
L3	Architecture-boundary compliance + commit format + lint/security (codebase-wide; no per-feature artifacts)	l3-quality-gate.yml (import-linter + sonar/snyk + commit-format)
L4	L3 + spec completeness, Gherkin structure, NFR coverage, FIRST + 7 Virtue scores, eval_criteria schema	govkit validate + quality-gate.yml + eval-gate.yml
L5	L4 + LLM quality (DeepEval), adversarial safety (Promptfoo), retrieval quality (RAGAS), guardrails config, multi-agent topology	govkit validate + deepeval-gate + promptfoo-gate + guardrails-check + multi-agent-gate

---

Getting Help

• Issues and feature requests: github.com/Accelerated-Innovation/governed-ai-delivery/issues
• Contributing: See CONTRIBUTING.md for development setup, project structure, and contribution guidelines
• Changelog: See CHANGELOG.md for release history
• CI Reference: See ci/README.md for what's enforced vs predicted in each CI gate

---

License

Copyright 2026 Accelerated Innovation

Licensed under the Apache License, Version 2.0. See LICENSE for details.

---

Glossary

Term	Definition
Agent	The AI coding tool (Claude Code, GitHub Copilot, or OpenAI Codex) that reads governance artifacts and generates code
Rule (Claude Code)	A path-scoped .md file in .claude/rules/ that loads automatically when editing files matching its path
Skill (Claude Code)	A reusable prompt in .claude/skills/ invoked via slash command (e.g., /architecture-preflight)
Instruction (Copilot)	A path-scoped .md file in .github/instructions/ — Copilot equivalent of a rule
Skill (Copilot)	A reusable task in .github/skills/ invoked via slash command — open agent skills standard
AGENTS.md (Codex)	A markdown instructions file read by Codex. A root AGENTS.md applies globally; nested AGENTS.md files at layer directories scope rules to that subtree via directory walk
Skill (Codex)	A SKILL.md under .agents/skills/<name>/ invoked via $skill-name
Marker (.govkit)	The marker directory/file written by govkit apply recording agent, level, type, stack, CI, assumptions, and calibration decisions so later commands auto-detect your configuration
govkit calibrate	Guided 9-step review that aligns the installed generic defaults with your actual repo and records the decisions in the marker
govkit doctor	Read-only governance-fit validator (rule globs, CI/stack/language match, stale baselines, extension manifests); complements govkit validate
govkit validate	CLI command that checks all features for governance compliance (artifact completeness, thresholds)
Port	An interface defining a contract between layers (inbound ports for API entry, outbound ports for infrastructure)
Adapter	An implementation of a port that connects to infrastructure (database, external API, message queue)
Domain	Business logic that has no dependencies on frameworks or infrastructure
FIRST	Test quality framework — Fast, Isolated, Repeatable, Self-verifying, Timely (scored 1–5)
7 Virtues	Code quality framework — Working, Unique, Simple, Clear, Easy, Developed, Brief (scored 1–5)
ADR	Architecture Decision Record — documents and gates architectural changes
NFR	Non-Functional Requirement — performance, security, availability, etc.
Evaluation Prediction	Predicted FIRST and Virtue scores in plan.md — must average >= 4.0 before implementation
Extension	An optional, self-describing pack under extensions/<id>/ that layers extra contracts/templates onto the core kit (drop-in; not CLI-installed)
/architecture-preflight	Agent skill that validates a feature against architecture contracts before planning ($architecture-preflight in Codex)
/genai-preflight	L5 agent skill that validates LLM gateway, observability, guardrails, and evaluation decisions ($genai-preflight in Codex)
/eval-suite-planning	L5 agent skill that plans DeepEval, Promptfoo, and RAGAS test suites ($eval-suite-planning in Codex)
LiteLLM	L5 sole LLM gateway — model routing, provider abstraction, fallback, cost tracking
OpenLLMetry	L5 LLM telemetry emission standard (OpenTelemetry for LLMs)
Langfuse	Trace storage, prompt versioning, and production evaluation visibility
DeepEval	L5 LLM quality evaluation — faithfulness, relevancy, hallucination metrics
Promptfoo	L5 adversarial and regression testing — jailbreak, injection, safety
RAGAS	L5 retrieval-specific evaluation — context recall, precision (RAG pipelines only)
NeMo Guardrails	L5 conversational safety — dialog flow control, topic boundaries, jailbreak prevention
Guardrails AI	L5 structured output validation — JSON schema enforcement on LLM responses