Embed AtomWorks Studio with chat-led RFD4 execution

.gitattributes

@@ -0,0 +1,3 @@
+# Generated bundle includes upstream shader/string whitespace verbatim.
+integrations/atomworks-studio/assets/studio.html -whitespace linguist-generated=true
+skills/bionemo-agent-toolkit/atomworks-studio/assets/studio.html -whitespace linguist-generated=true

.github/workflows/atomworks-studio.yml

@@ -0,0 +1,41 @@
+name: atomworks-studio
+on:
+  pull_request:
+    paths:
+      - 'integrations/atomworks-studio/**'
+      - 'tests/**'
+      - 'skills/bionemo-agent-toolkit/atomworks-studio/**'
+      - '.gitignore'
+      - '.github/workflows/atomworks-studio.yml'
+  push:
+    branches: [main]
+jobs:
+  adapter:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Verify Git-distributed Studio assets
+        run: python -m unittest discover -s tests -p test_studio_distribution.py -v
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '24'
+      - run: npm install --global pnpm@11.25.0
+      - run: pnpm install --frozen-lockfile --ignore-scripts && node --check host/bridge.js && node --check build-upstream.mjs
+        working-directory: integrations/atomworks-studio
+      - uses: actions/checkout@v4
+        with:
+          repository: RosettaCommons/atomworks-studio
+          ref: 4b419ebd327405035c8cf6ccde7f333d0878888e
+          path: .studio-upstream
+      - run: node test-public-import.mjs
+        working-directory: integrations/atomworks-studio
+        env:
+          ATOMWORKS_STUDIO_SOURCE: ${{ github.workspace }}/.studio-upstream
+      - run: node --test tests/test_atomworks*.mjs tests/test_studio_autoload.mjs
+      - run: pip install -r integrations/atomworks-studio/requirements.txt -r tests/requirements-rfd4-rosetta.txt
+      - run: python -m unittest discover -s tests
+      - name: Verify canonical RFD4 helpers
+        run: python -B -m pytest -q -p no:cacheprovider tests/test_rosetta_helpers.py tests/test_campaign_workflows.py tests/test_setup_preflight.py

.github/workflows/plugin-sync.yml

@@ -14,6 +14,7 @@ on:
       - "library-skills/**"
       - "open-models-skills/**"
       - "workflows/**"
+      - "integrations/**"
       - "skills/**"
       - "plugins"
       - "skills.sh.json"

.gitignore

@@ -9,3 +9,8 @@
 .venv/
 __pycache__/
 *.py[cod]
+
+# Frontend dependencies are reproducible from the pinned lockfile.
+node_modules/
+
+# Public Studio bundles and their provenance/licenses are tracked for Git plugin installs.

integrations/atomworks-studio/README.md

@@ -0,0 +1,151 @@
+# RFD4-Proteina Studio
+
+An embedded MCP App for BioNeMo Agent Toolkit. It packages AtomWorks Studio's
+existing molecular editor with a preloaded PD-L1 example, 3D annotations, saved
+drafts, reviewed RFD4 generation, results and chat-led execution on a named SSH host.
+Image notes accompany the input; they do not become model constraints automatically.
+The app uses a bundled frontend rather than an HTTP iframe. No NGS plugin is required.
+
+## Install and open
+
+Install the toolkit plugin from this repository's `skills/bionemo-agent-toolkit`
+directory. Its `.mcp.json` starts separate app and agent servers using `uv` and
+Python 3.12. The launcher installs the pinned MCP and CPU editor requirements
+through `uv`; the first launch needs network access and may take longer. Opening
+Studio automatically selects and saves a local loopback address and starts the
+CPU editor. No editor environment variables, GPU, model weights or remote compute
+registration are needed to view and annotate the preloaded example. Gradio is
+not required. Later launches reuse the saved editor address and session store.
+
+For custom installations, `RFD4_EDITOR_URL` or `editor.json` can select an existing
+compatible CPU editor. `RFD4_EDITOR_PYTHON` can select a separate editor environment
+prepared using `editor-requirements.txt`. `RFD4_STUDIO_PYTHON` bypasses `uv` and must
+provide `requirements.txt` plus `editor-requirements.txt` (unless a separate editor
+is configured). The editor runs locally and starts again on demand after a host
+restart. `RFD4_STUDIO_HOME` selects persistent state (default
+`~/.local/share/bionemo-rfd4`). Credentials stay in SSH/Brev configuration.
+
+Ask chat to open **RFD4-Proteina Studio**, then:
+
+1. Load the PD-L1 example, realize its editable segment and annotate residues.
+2. Choose thinking, design count, seed and optional sampling settings. Review
+   freezes the annotated CIF and settings without needing compute.
+3. Choose **Generate in chat**. Chat reads the saved request and asks for an SSH
+   host when the destination is unclear, verifies readiness, and submits it.
+4. Open the completed design from **Runs**. Use **File → Save CIF** or the run's
+   ZIP export. When Codex cannot download directly, chat retrieves the verified
+   local export and returns a file link.
+
+Drafts preserve annotations, notes and settings. Opening a result saves the current
+draft. The toolbar reports the run's machine and job state. Unsupported chat
+handoff preserves the request and explains how to continue; it never reports a
+successful submission.
+
+## Compute and runtime
+
+Compute is user-managed. The plugin checks a named SSH host, submits immutable
+requests, checks jobs and collects results/logs. It does not provision, start,
+stop or own machines. The `rfd4-studio-chat` skill guides host choice and setup.
+If Brev CLI is installed, chat can offer existing or new compute through that CLI.
+Creation requires a price and shutdown policy reviewed in chat. Existing machines
+stay running unless separately requested. No Brev installation, driver upgrade or
+reboot happens silently.
+
+SSH host and TLS verification stay enabled. Brev first-use trust can be approved
+in chat; changed keys are refused. Launch validation, upload and acknowledgement
+happen inside the active tool call. Only remote execution is detached. Durable
+receipts prevent duplicate inference after retries or interruption. Closing chat
+does not stop remote work or billing; resume in chat to collect and complete any
+authorized shutdown. Verified collected results remain available after stopping.
+
+Public model source is pinned to `RosettaCommons/RFD4-Proteina` at
+`c7c6cacbc617ab074b4e234448bed6581840949f`; the Hugging Face bundle is pinned at
+`a06e2509c7033b010fc43614e9fd840303588502`. `scripts/runtime_recipe.json` and
+`runtime_dependencies.lock` define the verified Python 3.12.10/Linux x86_64/CUDA
+13.2 runtime, AtomWorks 3.0.4 and TMol 0.1.64. Setup checks source, dependency and
+bundle hashes. Generation uses the local bundle in Hub offline mode. The public
+setup path requires no APT/compiler installation. Use a separate runtime root for
+this release; setup refuses to replace a different checkout revision. Model weights,
+private source, credentials, local state and evaluation bundles are not packaged.
+
+Fresh Brev images may have an older NVIDIA driver even when the machine has an
+A100. The preliminary setup check requires driver major version 580 or newer for
+this pinned runtime; passing that check alone does not establish readiness. If it
+fails, ask chat to inspect the selected machine's OS and `nvidia-smi` output.
+Choose a compatible provider image, or review a driver upgrade on that machine
+using NVIDIA's official packages. A driver upgrade may require a reboot and must
+wait until active work is finished. The plugin does not silently upgrade drivers,
+reboot, or create a replacement machine. After an approved upgrade/reboot,
+reconnect the same machine, repeat setup, and require all runtime checks plus the
+setup smoke test to pass before generating a design. The fresh A100 UAT image
+required a manual upgrade from 565.57.01 to 580.196.07; fresh provisioning is not
+qualified as an unattended workflow.
+
+Existing drafts, annotations, inputs and results are preserved in place. Legacy
+job records retain their original reader and are never replayed through the new
+submission path. Legacy compute records are suggestions only, not active selections.
+
+## Build and package
+
+Frontend source is public AtomWorks Studio v0.0.1, commit
+`4b419ebd327405035c8cf6ccde7f333d0878888e`. The CPU editor uses
+`atomworks-studio==0.0.1`. Narrow build-time adapters change transport, session
+coordination, storage and inline WASM without editing upstream source. Mol* is the
+renderer; Patinae, PWA installation and optional MP4 encoding are excluded.
+
+With Node 24, pnpm 11.25, Rust and wasm-pack, check out that Studio revision and
+run `pnpm install --frozen-lockfile` and `pnpm --dir packages/web build:core` there.
+Then, from this integration directory:
+
+```sh
+pnpm install --frozen-lockfile --ignore-scripts
+ATOMWORKS_STUDIO_SOURCE=/path/to/atomworks-studio pnpm build
+ATOMWORKS_STUDIO_SOURCE=/path/to/atomworks-studio node test-public-import.mjs
+RFD4_STUDIO_HOME=/path/to/dev-state RFD4_PREVIEW_PORT=8769 python scripts/preview.py
+```
+
+The tracked HTML, provenance, Studio license and dependency notices make Git
+installations self-contained. The installed payload includes only the server,
+Python helpers/profiles/locks, runtime requirements, this README and bundled
+assets/attribution. Frontend source, build tooling and the localhost preview stay
+in the development integration directory. From the repository root:
+
+```sh
+python scripts/plugin_sync.py --write
+python scripts/plugin_sync.py --check
+python scripts/stage_rfd4_plugin.py --output /absolute/path/to/new-plugin
+python -m unittest discover -s tests -p 'test_atomworks_*.py'
+python -m unittest discover -s tests -p test_studio_distribution.py
+```
+
+Canonical RFD4 skills retain their CC-BY-4.0 text and BSD-3-Clause helper attribution;
+see the root NOTICE and canonical skill README. The header embeds NVIDIA's official
+transparent green-eye/white-wordmark SVG; NVIDIA branding remains subject to trademark
+terms (source attribution in the development integration's `host/BRANDING.md`). Build provenance and third-party notices must
+remain with the bundled viewer.
+
+The Python adapter is organized by responsibility: `service.py` keeps editor,
+drafts and result operations; `design_request.py` freezes requests;
+`chat_execution.py` handles named-host submission and collection; `ssh_job.py`
+provides remote execution receipts. `registry.py` retains shared SSH transport,
+and `brev_provider.py` only reads provider identity for a named Brev target.
+There is no machine registry, provisioning service, ownership claim or local
+submission worker. Authored source/test files stay below 1,000 lines; generated
+viewer assets and dependency locks are retained intact.
+
+## Validation scope
+
+The chat-led path has completed one public PD-L1 design on the existing Brev
+A100, recovered it from a fresh client process, collected its results and logs,
+and opened the 1,454-atom result in the localhost editor. The existing machine
+remains running. This is source/localhost evidence, not an installed-plugin pass.
+
+The current release still requires an installed-plugin generation, chat handoff
+and export pass. A separate new A100 passed after an explicitly approved driver upgrade to
+580.196.07 and reboot: all nine runtime checks, one 866-atom short-monomer
+design, recovery from a fresh client, collection and offline ZIP export. The
+new machine was stopped after collection; the existing A100 stayed running. Earlier installed
+release results do not qualify this changed submission path. The broad runtime
+profile remains `candidate_not_qualified`; ChatGPT.com deployment is separately
+unqualified. Test evidence and outstanding checks are recorded in PR #58.
+Inference success is not scientific validation; use the canonical evaluation skills.

integrations/atomworks-studio/assets/STUDIO_LICENSE.md

@@ -0,0 +1,31 @@
+BSD 3-Clause License
+
+Copyright (c) 2026, Institute for Protein Design, University of Washington
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+Third-party dependencies and the Patinae patch series retain their respective
+licenses. See `docs/PROVENANCE.md` and `vendor/patinae/UPSTREAM.toml`.