OpenClinXR is a Step 2 CS–inspired XR clinical skills exam platform: timed multi-station encounters, WebXR learner runtime, faculty/admin review, scenario authoring, and an evidence-gated path from case definitions to runtime scenes.
Claim control: this is not an exam-equivalence product and does not claim clinical validity, licensure readiness, or production Quest certification. Early-stage R&D and local tooling first.
Project page:developers.simnova.com/OpenClinXR
Site source (GitHub Pages):docs/ on branch main
| Tier | What it is |
|---|---|
| Production exam platform | Author, review, assemble, run, trace, and replay timed clinical encounters (learners, faculty, admins, scenario authors). |
| Encounter Blueprint Factory | Turn reviewed case definitions into WebXR scenes, actor behavior, dialogue policies, emotion timelines, review packets, and persistence records. |
| Clinical Asset Commons | Reuse rooms, equipment, clothing, humanoids, animation, and provenance across encounters. |
| Capability Arena | Cage-match candidate tech (TTS, speech, humanoid generation, IWSDK sidecars, providers) before any production promotion. |
Local/offline is the default development path. Promotion gates for providers, Quest readiness, learner production, and clinical/scoring claims stay false unless explicitly approved and evidenced.
Pins live in mise.toml and package.jsonengines / packageManager. Prefer mise over nvm/system Python. Full detail: docs/TOOLING.md.
| Tool | Notes | |
|---|---|---|
| Must | mise | Project pins win; owns Node / pnpm / Python |
| Must | Node 24 (LTS) | engines.node: >=24.15.0 |
| Must | pnpm 11.18.x | packageManager: pnpm@11.18.0 — do not use corepack |
| Must | Python 3.13 via mise | Asset / Anny / voice scripts (python3) |
| Must | git | Clone and worktrees |
| Must | direnv (recommended) | Activates mise + loads .env.local on cd |
| Optional | Bun 1.3.x | Local experiments (mise pin) |
| Optional | MongoDB | Durable persistence; many smokes use in-memory fixtures |
| Optional | Blender | Anny / humanoid asset pipeline |
| Optional | GitHub CLI (gh) | PRs/issues — preferred over any GitHub MCP |
| Optional | DEEPSEEK_API_KEY | Agent assist only; never commit |
| Optional | Playwright browsers | pnpm playwright:install for browser evidence |
# Install mise: https://mise.jdx.dev# Install direnv: https://direnv.net
mkdir -p ~/.config/direnv/lib
mise direnv activate >~/.config/direnv/lib/use_mise.sh
# ~/.zshenv — shims first on PATH:# export PATH="$HOME/.local/share/mise/shims:$PATH"# ~/.zshrc:# eval "$(direnv hook zsh)"# Open a new terminal after editing shell config, or: exec zsh -lDo not combine eval "$(mise activate zsh)"and direnv use mise — pick direnv + shims (this repo’s pattern).
git clone <repo-url> openclinxr &&cd openclinxr
mise trust
mise install
cp .envrc.example .envrc
cp .env.local.example .env.local # optional: DEEPSEEK_API_KEY, MONGODB_URI, …
direnv allow
pnpm install.envrc is use mise only — secrets go in .env.local, never in .envrc.
Without direnv, from the repo root:
eval"$(mise env -s zsh)"# or: mise activate zshpnpm env:doctor # mise pins, PATH, turbo, install health# or: mise run doctor
pnpm openclaw:preflight # env:doctor + alignment + drift + lease status
pnpm local:exam:smoke # deterministic ED station harness (no cloud)Useful variants: pnpm env:doctor:json, pnpm env:doctor:strict.
Each in its own terminal, from repo root with mise/direnv active:
pnpm --filter @openclinxr/api dev
pnpm --filter @openclinxr/ui-admin dev:portless # default port 5174
pnpm --filter @openclinxr/ui-xr dev:portless # default port 5173Arena / tooling (optional):
pnpm --filter @openclinxr/model-vetting-studio dev:portless # humanoid cage-match UI
pnpm arena:iwsdk:dev # IWSDK WebXR spike| Path | Role |
|---|---|
apps/api, apps/ui-admin, apps/ui-xr | Production-facing API and UIs |
apps/arena/* | Capability Arena sidecars (IWSDK, voice, model vetting, …) |
packages/openclinxr/* | Domain, runtime, gateways, persistence, review, fixtures |
packages/openclinxr/arena/* | Spike / experimental packages |
tools/openclinxr/factory/* | Blueprint → runtime / review / materialization generators |
tools/openclinxr/evidence/* | Validators, benchmarks, capture helpers |
tools/openclinxr/openclaw/* | Build-ops CLIs (env doctor, lease, preflight, …) |
docs/openclinxr/ | Product docs, runbooks, evidence |
docs/TOOLING.md | Toolchain, PATH, agent shells |
Encounter flows are meant to be blueprint-driven through the factory, not hand-authored one-off scenes.
OpenClinXR ships generated 3D assets. Several tools that can produce them carry licences whose terms depend on facts about the product, not about the code — so the same tool can be fine today and prohibited later. The parameters are recorded in MADR 0046:
| Parameter | Current value |
|---|---|
| Annual revenue | below USD 1,000,000 |
| Operating territory | United States only |
| Distribution model | not yet stated — affects notice obligations |
If either of the first two changes, re-read this table before generating another asset. Two of the restrictions below terminate or reach backwards into assets already produced.
| Tool | Gate | What happens at the threshold |
|---|---|---|
| Stable Fast 3D | Stability AI Community License, ~USD 1M annual revenue (org-wide, any source) | The licence terminates automatically. An Enterprise licence is required to continue. Commercial use below the threshold also requires registration at stability.ai/community-license. |
Treat the revenue threshold as a hard product gate with a tripwire, not a reminder. Outputs you already generated are yours (§IV.c.iii), but continued use of the model is not.
| Tool | Gate | What happens outside it |
|---|---|---|
| Hunyuan3D-2 | Tencent Community License: Territory is worldwide excluding the EU, the UK and South Korea | AUP §5.c bars distributing or displaying the Outputs outside Territory — not just the model. Expanding to those markets re-blocks meshes already generated. |
Because the restriction follows the Output, anything generated this way must record its generator in the MADR 0016 asset manifest, not merely its licence — otherwise a future territory change means auditing every asset by hand.
Hunyuan also carries a 1M-monthly-active-user gate measured at its 2025-01-21 release date, and an AUP whose medicine clauses (§14 high-stakes automated decisions, §20 unlicensed practice) restrict automated clinical decision-making, not the generation of props for a training simulator. Neither binds today.
| Tool | Licence | Posture |
|---|---|---|
| MPFB2 | GPL-3.0-or-later | Out-of-repo authoring tool. Never vendored, never imported by repo code, never shipped. |
| StableGen | GPL-3.0 | Same. Also does not complete headless — its generation is a modal Blender operator. |
| ComfyUI | GPL-3.0 | Same. Output textures are governed by the checkpoint licence, not ComfyUI's. |
The outputs of these tools ship; their code does not. That split is the whole posture — see MADR 0044.
For MakeHuman-derived assets the safer factual basis is the asset licence, not the tool's: the
hm08 base mesh is explicitly CC0, and each community garment carries its own license: field in its
.mhclo header — machine-readable, so a parser can feed the manifest directly.
| Tool | Why |
|---|---|
| TRELLIS | Core is MIT, but diffoctreerast and the Inria gaussian-splatting family are research / non-commercial. Also requires CUDA, which this workstation does not have. |
| Infinigen (whole rooms) | BSD-3-Clause and licence-clean. Blocked on geometry, not licence: ~15.5M triangles against a 180,000 per-station budget. Its infinigen_gpl submodule is snow and particles, and is not on the indoor path. |
Poly Haven and ambientCG are CC0 by site policy — but their APIs carry no per-asset licence field, so record the policy URL plus a byte hash in the manifest rather than scraping a page. NIH 3D is per-entry and must be read individually.
This section is an engineering reading of licence texts, not legal advice. Counsel should sign anything that ships.
| Doc | Role |
|---|---|
| PROJECT_STATUS.md | Canonical project state (priority, active work, backlog, strategy) |
| docs/openclinxr/worker-backlog-and-validation-matrix.md | Worker ownership + validation matrix |
| docs/TOOLING.md | mise, direnv, env doctor, MCP→CLI |
| AGENTS.md | Agent operating contract (for AI/agent contributors) |
| docs/openclinxr/ | Product docs and evidence |
| docs/madr/README.md | Architecture decision records / arena-to-decision map |
| docs/madr/0046-... | Licence gate parameters — revenue band, territory, and what they unblock |
Not live SSOT (historical names only — recover from git history if needed; do not recreate as living ledgers):
AUTONOMOUS_WORK_PLAN.md/PROJECT_COORDINATION_INDEX.md→ purged 2026-08-05; successor isPROJECT_STATUS.md
Warehouse process: docs/agent-ops/DOC-WAREHOUSE.md. Status purge audit: docs/openclinxr/reviews/2026-08-05-status-doc-purge-manifest.md.
OpenClaw is the repo-native build operating model, not the product. There is no external OpenClaw daemon or SaaS required to develop.
Work is sliced, lease-gated, and checked with deterministic scripts so humans and coding agents share the same guardrails. Day-to-day for developers:
pnpm env:doctor # “can I run tools?”
pnpm openclaw:preflight # broader readiness (env + docs alignment + drift + lease)
pnpm local:exam:smoke # cheap product-path smoke without cloud servicesAgents and long-running automation also use pnpm openclaw:post-slice, leases, and the protected runbooks under docs/openclinxr/. Contributors who only need apps and packages can stay on the Get started path above and ignore agent automation until they need it.
Condensed host prompts (full contract: AGENTS.md). State SSOT is PROJECT_STATUS.md (not archived coordination ledgers).
Codex
Continue in repo-native OpenClaw mode in /Volumes/files/src/openclinxr using Codex local tools.
Read AGENTS.md, PROJECT_STATUS.md (snapshot), docs/openclinxr/worker-backlog-and-validation-matrix.md, docs/openclinxr/openclaw-runbook-2026-05-27.md, docs/openclinxr/openclaw-tool-adapters-2026-05-27.md, and docs/openclinxr/blueprint-factory-drift-guardrails-2026-05-27.md as needed.
Use terminal, file edits, focused verification. Run pnpm docs:drift-check and pnpm agent:alignment before long unattended work. Select the next approved product slice from PROJECT_STATUS.md Next dequeue and continue without treating slice completion as a stop condition.
Claude
Operate as a repo-native OpenClaw agent for /Volumes/files/src/openclinxr, not as generic Claude chat.
Use AGENTS.md, PROJECT_STATUS.md, docs/openclinxr/openclaw-runbook-2026-05-27.md, and docs/openclinxr/openclaw-tool-adapters-2026-05-27.md as the source of truth. Keep work blueprint/factory-driven.
If you have shell and file access, implement the next smallest approved product slice and run focused verification. If not, act as a bounded planner/reviewer without inventing status ledgers.
Grok
Main session = orchestrator only (chief-coordinator). Spawn role-mapped subagents for product IC.
Use AGENTS.md BLUF, PROJECT_STATUS.md snapshot, docs/agent-ops/PATH-SCOPE.md / TEMPORAL-DECISIONS.md / TASK-COST-ROLLUP.md as needed.
Prefer pnpm openclaw:slice-token:start → team work → openclaw:slice-token:finish for token + task cost lines.
Cursor
Run Cursor in repo-native OpenClaw mode for /Volumes/files/src/openclinxr.
Use AGENTS.md, PROJECT_STATUS.md, docs/openclinxr/openclaw-runbook-2026-05-27.md, and blueprint-factory-drift-guardrails before editing.
Make focused diffs against the next approved product slice; run the smallest relevant verification before claiming completion.
pnpm agent:alignment # cheap coordination check
pnpm docs:drift-check # doc / coordination drift
pnpm pages:validate # public site (docs/) consistency# Focused package test example:
pnpm --filter @openclinxr/api test -- app.test.ts -t "name"Prefer focused filters over full monorepo pnpm verify unless you are closing a release-style gate. See docs/TOOLING.md for turbo agent vs human scripts and known tooling constraints.