medharness 0.3.7

Open-source design-controlled development infrastructure for medical device and SaMD teams — includes dhfkit

MedHarness

AI harness and DHF tooling for medical device software teams.

MedHarness structures how AI agents interact with a Design History File under IEC 62304 / FDA-regulated software projects. It pre-computes DHF context before an agent runs, enforces approval gates the agent must pass through, and commits decisions back into the DHF — so the engineer controls the feedback loop, not the agent.

It combines two packages:

• medharness — CLI harness, CI gates, CR workflows, project scaffolding (init)
• dhfkit — standalone DHF engine for items, traceability, document generation, schema validation

---

Install

pip install medharness[full]

[full] pulls in optional extras: ai (Gemini-based AI review) and docs (PDF export via WeasyPrint). Omit for a minimal install — the DHF engine (dhfkit) is always included.

Verify:

medharness --help
dhfkit --help

From source (development):

git clone https://github.com/itercharles/MedHarness
cd MedHarness
pip install -e ".[dev]"
pytest dhfkit/tests/ tests/

---

Quick Start

medharness init is zero-prompt — it scaffolds a single-repo project in the current directory. The project name is derived from the directory name.

mkdir my-medical-device && cd my-medical-device
python -m venv .venv && source .venv/bin/activate
pip install medharness
medharness init

After init completes, here's what exists on disk:

my-medical-device/                  # single repo — DHF + source together
├── DHF/
│   ├── config/
│   │   ├── global.yaml             # project name, lifecycle states
│   │   └── doc_types/              # one YAML per type (SYS, CRS, SRS, SWDD, CR, …)
│   ├── items/                      # one YAML file per requirement / risk / CR
│   │   ├── 01_crs/                 # Customer Requirements (CRS-NNN.yaml)
│   │   ├── 02_sys/                 # System Requirements (SYS-NNN.yaml)
│   │   ├── 03_srs/                 # Software Requirements (SRS-NNN.yaml)
│   │   ├── 06_cr/                  # Change Requests (CR-NNN.yaml)
│   │   └── ...                     # Use Cases, SOUP, Risk, Defects, etc.
│   ├── test-results/
│   ├── documents/
│   │   ├── specs/                  # Jinja2 spec templates (.j2)
│   │   └── plans/                  # development_plan.md, verification_plan.md, …
│   └── README.md
├── .github/
│   └── prompts/                    # optional prompt files for repo-local automation
├── tests/                          # product test suite
├── CLAUDE.md                       # agent entrypoint
├── .gitignore
└── README.md                       # project README

The scaffolded items are starter samples — replace them with your project's real requirements, architecture, and plans before using this for a regulated product.

Initialize git and push:

git init && git add -A
git commit -m "feat: initialize My Medical Device with MedHarness"
git remote add origin https://github.com/<org>/my-medical-device
git push -u origin main

---

Automation Model

MedHarness no longer ships prescribed GitHub workflow files as part of the product surface. The stable interface is the CLI.

Use the CLI directly from whichever automation layer you prefer:

• GitHub Actions
• GitLab CI
• Jenkins
• local scripts
• internal orchestration systems

Typical entrypoints are:

medharness ci dhf-validate --dhf DHF
medharness ci test-coverage --dhf DHF --junit-dir test-results
medharness --dhf DHF ci analyze-cr --cr CR-034
medharness --dhf DHF ci design-cr --cr CR-034
medharness --dhf DHF ci develop-cr --cr CR-034
medharness --dhf DHF ci validate-design --cr CR-034
medharness --dhf DHF ci validate-code --cr CR-034
medharness --dhf DHF ci validate-branch --cr CR-034 --code-path src/
medharness ci cr-status --cr CR-034 --stage spec --pr 18
medharness --dhf DHF ci evidence bundle --out-dir artifacts --junit-dir test-results
medharness ci github-event --event "$GITHUB_EVENT_PATH"

How a Change Request flows

Every non-trivial change starts as a Change Request (CR) in the DHF. CRs move through AI-assisted stages, each gated by human approval. How those stages are wired into automation is up to the client repo:

Issue → CR review → analyze-cr → design-cr → develop-cr → cr-complete

Stage	Trigger	What MedHarness does
CR intake	Issue milestoned	Creates CR item in DHF, opens draft PR (cr workflow intake-github-issue-ci)
analyze-cr	CR PR approved	Runs Claude to write a spec, self-corrects against schema, commits to docs/cr-specs/ (ci analyze-cr)
design-cr	Spec PR approved	Runs Claude to create/update DHF items, validates schema + traceability (ci design-cr)
develop-cr	Design PR approved	Runs Claude to implement code, opens implementation PR (ci develop-cr)
cr-complete	PR merged	Transitions CR to completed in the DHF (cr workflow complete-from-github-pr)

When a PR receives review feedback, re-run the same command with --pr N to revise the existing output based on reviewer comments.

To let external automation decide whether a CR stage is ready to advance, use the CLI's machine-readable status surface rather than embedding policy in workflow YAML:

medharness ci cr-status --cr CR-034 --branch spec/CR-034 --pr 18

To catch deterministic issues before a PR is opened, client automation can run the same preflight validators directly:

medharness --dhf DHF ci validate-design --cr CR-034
medharness --dhf DHF ci validate-code --cr CR-034 --since-ref origin/main
medharness --dhf DHF ci validate-branch --cr CR-034 --since-ref origin/main --code-path src/

validate-branch always checks that the approved spec file exists and that DHF item changes are present when the spec expects them. Pass --code-path <dir> (may be repeated) to also require implementation code changes under those paths. Omitting --code-path skips the code-change check — use this when a CR touches only documentation or DHF items with no product-code impact.

---

Test Coverage Gate

The CI gate (medharness ci test-coverage) enforces that every verifiable requirement has at least one passing test linked to it.

JUnit XML contract

Tests must emit JUnit XML with properties linking to DHF item IDs:

<testcase name="test_TC_SYS_005_001_validates_link_format">
  <properties>
    <property name="medharness.id" value="TC-SYS-005-001"/>
    <property name="medharness.links" value="SYS-005"/>
  </properties>
</testcase>

All property names are defined as constants in medharness/contracts.py:

Property	Purpose
medharness.id	Test case identifier (e.g. TC-SYS-005-001)
medharness.links	Comma-separated DHF item IDs the test covers
medharness.title	Human-readable test title (optional)
medharness.reviewer	Reviewer name (optional)
medharness.review_date	Review date (optional)
medharness.review_status	Review status (optional)

Python / pytest

Use pytest's record_property in conftest.py:

@pytest.fixture(autouse=True)
def _inject_medharness_metadata(request, record_property):
    doc = request.function.__doc__ or ""
    tc_id = extract_tc_id_from_name(request.node.name)
    links = parse_links(doc)   # extract @links:SYS-005 from docstring
    if tc_id:
        record_property("medharness.id", tc_id)
    if links:
        record_property("medharness.links", ",".join(links))

TypeScript / Vitest / Playwright

Use custom JUnit reporters that emit <properties> blocks for medharness.links. Reference implementations are available in the WebTPS repo.

Running the gate locally

# From project root
pytest tests/ -q --junitxml=test-results/results.xml
medharness --dhf DHF ci test-coverage --junit-dir test-results

Expect output like:

[test-coverage] SRS: 12/14 covered
      ↳ uncovered: SRS-012
      ↳ uncovered: SRS-008

The command exits non-zero when gaps exist, blocking CI.

---

CLI Reference

Scaffold

medharness init                     # zero-prompt single-repo project setup

DHF operations (run with --dhf DHF)

medharness --dhf DHF dhf item list --type SYS
medharness --dhf DHF dhf item get SYS-001
medharness --dhf DHF dhf item create --type SYS --data '{"title": "My req"}'
medharness --dhf DHF dhf item update SYS-001 --data '{"title": "Updated"}'
medharness --dhf DHF dhf item delete SYS-001
medharness --dhf DHF dhf item transitions CR-001
medharness --dhf DHF dhf item transition CR-001 approved --by "Alice"
medharness --dhf DHF dhf validate schema
medharness --dhf DHF dhf validate traceability
medharness --dhf DHF dhf doc list
medharness --dhf DHF dhf doc generate SYS
medharness --dhf DHF dhf doc export SYS          # PDF output (requires `[docs]`)
medharness --dhf DHF dhf test list
medharness --dhf DHF dhf config doc-types

CI gates

medharness ci dhf-validate --dhf DHF
medharness ci test-coverage --dhf DHF --junit-dir test-results
medharness ci evidence bundle --dhf DHF --out-dir artifacts

CR generation commands

Encapsulate the full AI loop for each CR stage: prompt assembly (with embedded DHF impact skills) → claude -p invocation → validate → self-correct.

# Initial generation
medharness --dhf DHF ci analyze-cr --cr CR-034   # write docs/cr-specs/CR-034-Spec.md
medharness --dhf DHF ci design-cr  --cr CR-034   # create/update DHF items
medharness --dhf DHF ci develop-cr --cr CR-034   # implement code

# Revision based on PR review feedback
medharness --dhf DHF ci analyze-cr --cr CR-034 --pr 42
medharness --dhf DHF ci design-cr  --cr CR-034 --pr 42
medharness --dhf DHF ci develop-cr --cr CR-034 --pr 42

ANTHROPIC_MODEL env var selects the Claude model. GH_TOKEN is required when --pr is used (fetches review comments from the GitHub API).

Each command outputs JSON to stdout:

{ "cr_id": "CR-034", "stage": "spec", "status": "ok",
  "corrections": 0, "validation": "passed", "errors": [],
  "spec_path": "docs/cr-specs/CR-034-Spec.md",
  "spec_json_path": "docs/cr-specs/CR-034-Spec.json",
  "started_at": "2026-05-11T14:23:45+00:00", "elapsed_ms": 28500 }

CR workflow commands

medharness cr workflow intake-github-issue-ci      # CR intake from issue
medharness cr workflow complete-from-github-pr     # CR completion on PR merge

Agent session helpers

medharness ci claude-session put <pr_number> <session_id>
medharness ci claude-session get <pr_number>

---

Python API

Use DHFClient for high-level operations (recommended for product repo automation):

from medharness.client import DHFClient

client = DHFClient(Path("DHF"))

cr   = client.get_item("CR-034")
spec = client.get_cr_context("CR-034")   # {"cr": {...}, "spec": "..."}
client.transition_item("CR-034", "in_review", performed_by="alice")

Or use dhfkit standalone (no dependency on medharness):

from dhfkit.local_adapter import LocalDHFAdapter

adapter = LocalDHFAdapter(Path("DHF"))
items  = adapter.list_items("SRS")

---

Repository layout

Directory	Purpose
medharness/	CLI harness, CI gates, CR workflows, init scaffolding
dhfkit/	DHF engine: items, lifecycle, traceability, document generation
dhfkit/templates/	Starter DHF scaffold — config, specs, plans, sample items
tests/	MedHarness and dhfkit test suites
docs/	Architecture, ADRs, compatibility contracts

dhfkit has no dependency on medharness — the engine can be used standalone.

---

Docs

• docs/architecture.md — packages, scaffold model, DHF lifecycle
• docs/compatibility-contracts.md — stable public contracts
• docs/adr/ — architecture decision records
• CHANGELOG.md — version history

---

License

MIT — see LICENSE.