GL

glacius-labs/markitect

Developer tools
0 stars Quality 57 Trend 20Apache-2.0

Typed engineering knowledge for AI workflows: explicit ownership, dependencies, context, and change impact.

Overview

Markitect is model-first development with delegated realization. You maintain one canonical YAML model of your project's intended world under .markitect/: its concepts, rules, boundaries, responsibilities, checks and dependencies. Source code, tests, configuration and prose remain ordinary files in the adopting project; they realize the model. You change the model, usually in conversation with your coding agent, and Markitect has the change applied. It compiles the model, derives the affected Managers and files, and lets recursive Managers implement the change with independent review, integration, fixed-candidate verification and guarded Apply. Markitect reports evidence for the checks performed; project owners retain decisions about semantic acceptance and release. Markitect in brief explains why and how to judge it. The model-first markitect workflow is available in this development source checkout; it is not included in the published v0.14.1 binary.

README

Markitect

Markitect is model-first development with delegated realization. You maintain one canonical YAML model of your project’s intended world under .markitect/: its concepts, rules, boundaries, responsibilities, checks and dependencies. Source code, tests, configuration and prose remain ordinary files in the adopting project; they realize the model.

You change the model, usually in conversation with your coding agent, and Markitect has the change applied. It compiles the model, derives the affected Managers and files, and lets recursive Managers implement the change with independent review, integration, fixed-candidate verification and guarded Apply. Markitect reports evidence for the checks performed; project owners retain decisions about semantic acceptance and release. Markitect in brief explains why and how to judge it.

Quick navigation

Goal Start here
Build the model-first workflow from this source checkout Run the current source
Create a new adopting project Start a model-first project
Model an existing repository Brownfield workflow
Understand the workflow and tool boundaries Project workflow, Project operations, and Architecture
Install the published compatibility CLI Install Markitect and distribution guidance

Run the current source

The model-first markitect workflow is available in this development source checkout; it is not included in the published v0.14.1 binary. Building from source does not publish a release or change a project’s existing version pin.

For Windows, use Git and Go 1.27.1 or later. From this Markitect source checkout, build a separate development executable:

$toolBin = Join-Path $env:LOCALAPPDATA 'Markitect\development'
New-Item -ItemType Directory -Force -Path $toolBin | Out-Null
go build -o (Join-Path $toolBin 'markitect.exe') ./src/cmd/markitect
if ($LASTEXITCODE -ne 0) { throw 'Markitect build failed.' }
$env:PATH = "$toolBin;$env:PATH"
markitect help

Keep the source checkout for development. Configure the adopting project’s runtime with config, then connect the outer coding client to its repository-local Markitect MCP server. Manager, reviewer and verifier roles use the supported Codex App Server runtime by default. Setup can give each role class its own provider, model, effort and cost mode, including a bring-your-own process executor; an executor that Markitect cannot price runs unmetered, and its cost is reported as unknown, not zero. See Provider adapters for connection, role profiles and execution boundaries.

Start a model-first project

Initialize your target Git repository on a feature branch, then create its scaffold:

# Run in the target project, with the development executable on PATH.
git init -b codex/project-start
markitect init --repo . --name my-project
markitect init --repo . --name my-project --expect INIT_PLAN_DIGEST --write
markitect onboard --repo . --provider codex

Review the onboarding preview and apply it with the returned digest:

markitect onboard --repo . --provider codex --expect ONBOARDING_PLAN_DIGEST --write

Choose claude or both for those contributors. Onboarding previews repository-local entrypoints and discoverable skills, preserves custom content, and points contributors to the shared workflow. It does not register MCP or change global account/provider settings. Connect the selected client using Provider adapters.

Then give your agent an ordinary request, for example:

Build a small shop where creating an order reserves stock and cancelling it restores the reserved stock exactly once. Include tests and explain the project through its generated documentation.

Use the Project workflow to inspect the model, record scope and acceptance, update canonical intent when needed, establish readiness, and continue the same operation through execution, verification, and guarded Apply. The Project operations reference documents exact interfaces and recovery. The Shop example provides a runnable model and finite tests.

Example adopting-project layout

The tree below illustrates a repository that uses Markitect. It describes a target adopting project, not this Markitect source repository. In this source checkout, implementation is under src/, product documentation is under docs/, and adopting-project examples are under examples/.

.markitect/
  project.yaml                    project selection and policy
  model/
    manager.yaml                  root responsibility
    sales/
      manager.yaml                Sales responsibility
      orders/                     concepts, rules and use cases together
  runtime.yaml                    pinned local execution and budgets
  workflows/model-first.md        installed contributor workflow
  state/                          exploration and model-change history
  runs/                           candidates, receipts and verification
.agents/skills/                   native Codex skills, when selected
.claude/skills/                   native Claude skills, when selected
AGENTS.md / CLAUDE.md             selected native entrypoints
src/                             application artifacts in the adopting project
tests/                           tests in the adopting project
docs/markitect/project.md         generated readable model view

Application paths are choices made by each project, not imposed language conventions. The model tree follows conceptual slices. Each file has one accountable owner and explicit relationships to the concepts or rules it realizes; a shared file can realize several concepts. Managers receive bounded context and public neighbor contracts. Parents receive child results for integration, not private transcripts.

Change, verify, and recover

Start an intent change in the model. Review its readable view and commit the accepted canonical model under the repository’s policy before implementing it. Drafts remain proposals. Markitect records accepted first-parent model transitions and supplies scoped change briefings.

Implementation follows impact through responsible Managers and their ancestors. Leaf implementers and independent reviewers iterate within finite limits; parent Managers integrate child results and can request bounded rework. Final verification evaluates permanent Manager responsibilities and declared checks against one immutable candidate. Coverage accounts for ordinary files, explicit ignores, and known tool files so an unknown file cannot silently disappear.

The MCP-led workflow advances a ready named scope through planning, Manager work and integration, verification, preflight, and guarded Apply. A successful Apply records technical completion, not semantic acceptance. Interrupted work resumes with the same durable IDs; unknown outcomes are inspected rather than replayed. See Project operations for recovery and exact interfaces.

For an existing repository, begin with fixed-source discovery and iterative reverse modeling. Separate observed code behavior from documented intent and future decisions. Managers make bounded proposals; parent Managers integrate them, and explicit owner resolutions precede model-only adoption. Commit the accepted model before a separate cleanup or implementation request. Unresolved or transitional areas remain visible. See the Brownfield workflow.

Native instructions guide cooperation but do not prevent a process with ordinary filesystem permissions from editing files directly. Guarded Apply, checks, and repository policy enforce their stated boundaries. Commits, digests, and decision references bind evidence but do not authenticate human approval.

Current source status and compatibility

The combined A01 native smoke passed for source cdd30b0efc540f151404dabe86b022275dc40d83. That result is tied to its recorded source and does not establish human semantic acceptance or productivity benefit. The delivery plan owns the rebuild scope and coordination; chats/PRs record candidate evidence; the A01 validation record records the run and its limits.

Earlier Project/Domain contracts, released examples and historical research are preserved at their original immutable revisions, including the v0.14.1 compatibility guide. Existing projects need an explicit migration; the model-first workflow does not silently reinterpret old files. Use the distribution guide to maintain a released installation.

Documentation and development

Use the documentation map to find product and development guidance. The architecture describes product boundaries; contribution checks describes source validation. Adopting projects own their business rules, technologies, and acceptance policy.

License

Markitect is licensed under Apache-2.0. The Glacius Labs and Markitect names and logos are not licensed as trademarks by that license. Third-party notices are maintained with the distribution guidance.

View this README on GitHub

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add glacius-labs/markitect