MedHarness

AI-assisted development with design control: your AI agent writes the code and the design record, and ordinary code checks that they agree before anything merges.

AI development

flowchart LR
    A["<b>1 · Design</b><br/>AI drafts the design"] -->|"<b>you review</b>"| B["<b>2 · Code</b><br/>AI codes and tests"]
    B -->|"<b>you review</b>"| C(["Merge"])

For each change request, the agent designs first — it assesses the change's risk, then writes the requirements, design and test points it needs as items in your repository's Design History File (DHF), each traced to the next — and codes second, tagging each test with the requirement it verifies. You review the design, then the code. Between them, ordinary code — never a model — checks that the design traces, every requirement is verified, and the branch changed what the request said.

What MedHarness gives the agent:

• The process, in its own instructions. init writes an AGENTS.md (and a CLAUDE.md that imports it) telling the agent when to open a change request and which commands to run; build plan --prompt prints the exact steps and this change's DHF context for whichever agent you use.
• Commands instead of files. The agent reads and writes the DHF with medharness item, which checks the schema first: a bad edit is refused, not saved.
• Feedback it can act on. The checks name the item and field that are wrong, so the agent fixes them and runs the check again.

How the agent and MedHarness fit together — what to call when, a skill, and two Claude Code hooks — is in Working with your coding agent.

Run it two ways:

• At your desk, with the agent you already use: it does the work in your working tree and you commit.
• Unattended, from an issue: CI opens the change request and a pull request with the design; ask for changes and it revises; the code stage runs the same way. The recipe is in the AI workflow page.

Unattended, build plan and build code run the claude CLI, or any provider:model you set (anthropic, openai, deepseek, or a local OpenAI-compatible endpoint such as ollama) — they run an agent with a shell, so use an ephemeral runner and read ai-security.md first.

Quick start

pip install medharness
mkdir my-device && cd my-device
medharness init            # writes DHF/ with sample items, and AGENTS.md for your agent

Then ask your coding agent: "open a CR for PDF export and implement it". Or use the checks alone, with no AI: medharness verify dhf checks that the design holds together.

What a project looks like

my-device/
├── DHF/
│   ├── config/global.yaml       # the project name — and only what you change
│   └── items/                   # one YAML file per item, one directory per type
│       ├── 01_crs/CRS-001.yaml
│       ├── 03_srs/SRS-001.yaml
│       └── …                    # one directory per configured type
├── AGENTS.md                    # product context and DHF steps, for any coding agent
└── CLAUDE.md                    # @AGENTS.md, so Claude Code reads it too

The item types (13 by default), their fields and lifecycles, the required links and the specification templates are defaults shipped in the package, so upgrading medharness upgrades them. A project overrides only what it changes, in DHF/config/: see Changing the defaults.

An item is a small YAML file. Links are written on the child and point up:

id: SRS-012
title: Password must be at least 12 characters
derives_from: [SYS-004]
verification_method: [Test]
testing: |
  T1: an 11-character password is rejected

Commands

One CLI, four groups:

medharness init
medharness item    list | get | create | update | transition
medharness verify  dhf | tests | soup | completion | impact | changes
medharness build   plan | code | soup | release

item reads and changes DHF items, verify checks and writes nothing, build produces items, code and release artifacts. Every command's options and what it returns are in interface.md; --help on any command prints the same.

Example project

ContourLab is an example project used to exercise MedHarness end to end: its DHF, its CI, and changes made through AI development.

Documentation

License

MIT. See LICENSE.