inferport 0.1.0__tar.gz

inferport-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,66 @@
+name: InferPort
+
+on:
+  push:
+  pull_request:
+  workflow_call:
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python: ['3.10', '3.11', '3.12', '3.13', '3.14']
+        dependencies: [latest]
+        include:
+          - python: '3.10'
+            dependencies: minimum
+          - python: '3.11'
+            dependencies: numpy123
+          - python: '3.12'
+            dependencies: numpy126
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v6
+      - run: uv sync --python '${{ matrix.python }}' --group dev
+      - if: matrix.dependencies == 'minimum'
+        run: uv pip install numpy==1.21.3 msgpack==1.1.0 websockets==16.1.1
+      - if: matrix.dependencies == 'numpy123'
+        run: uv pip install numpy==1.23.5
+      - if: matrix.dependencies == 'numpy126'
+        run: uv pip install numpy==1.26.4
+      - if: matrix.dependencies == 'latest'
+        run: uv pip install --upgrade 'numpy>=1.21.3,<3' 'msgpack>=1.1,<2' 'websockets>=16.1.1,<18'
+      - run: uv pip check
+      - run: uv run --no-sync pytest -q
+      - run: uv run --no-sync ruff check .
+      - run: uv run --no-sync ruff format --check .
+      - run: uv build
+
+  cross-environment:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v6
+      - run: |
+          uv build
+          uv venv --python 3.10 .venv-310
+          uv venv --python 3.12 .venv-312
+          uv pip install --python .venv-310/bin/python dist/*.whl numpy==1.21.3 msgpack==1.1.0 websockets==16.1.1
+          uv pip install --python .venv-312/bin/python dist/*.whl 'numpy>=2,<3'
+          uv run --no-project --python .venv-312/bin/python tests/cross_environment.py --python-a .venv-310/bin/python --python-b .venv-312/bin/python
+
+  platform-smoke:
+    strategy:
+      matrix:
+        os: [windows-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v6
+      - run: uv sync --python 3.12 --group dev
+      - run: uv run pytest -q tests/test_codec.py tests/test_protocol.py tests/test_lifecycle.py tests/test_adapter_contracts.py tests/test_packaging.py

inferport-0.1.0/.github/workflows/publish.yml

@@ -0,0 +1,72 @@
+name: Publish to PyPI
+
+on:
+  workflow_dispatch:
+  release:
+    types: [published]
+
+permissions:
+  contents: read
+
+concurrency:
+  group: pypi-${{ github.event_name }}-${{ github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  verify:
+    uses: ./.github/workflows/ci.yml
+
+  build:
+    needs: verify
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+      - uses: astral-sh/setup-uv@v6
+      - run: uv sync --locked --python 3.12 --group dev
+      - name: Check release tag matches package version
+        if: github.event_name == 'release'
+        env:
+          RELEASE_TAG: ${{ github.event.release.tag_name }}
+        run: |
+          uv run --no-sync python - <<'PY'
+          import os
+          import inferport
+
+          expected = f"v{inferport.__version__}"
+          if os.environ["RELEASE_TAG"] != expected:
+              raise SystemExit(f"Release tag must be {expected}")
+          PY
+      - name: Build distributions and validate metadata
+        run: |
+          uv build --no-sources
+          uvx twine check --strict dist/*
+      - name: Test the built wheel in a fresh environment
+        run: |
+          uv venv --python 3.12 .venv-dist
+          uv pip install --python .venv-dist/bin/python dist/*.whl 'pytest>=8,<10'
+          uv pip check --python .venv-dist/bin/python
+          uv run --no-project --python .venv-dist/bin/python -m pytest -q
+      - uses: actions/upload-artifact@v5
+        with:
+          name: python-package-distributions
+          path: dist/
+          if-no-files-found: error
+
+  publish:
+    if: github.event_name == 'release' && github.repository == 'jeremy775885/InferPort'
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/p/inferport
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v6
+        with:
+          name: python-package-distributions
+          path: dist/
+      - name: Publish using the PyPI trusted publisher
+        uses: pypa/gh-action-pypi-publish@release/v1

inferport-0.1.0/.gitignore

@@ -0,0 +1,10 @@
+__pycache__/
+*.egg-info/
+dist/
+build/
+*.pyc
+.venv/
+.idea/
+.pytest_cache/
+.ruff_cache/
+.venv-*/

inferport-0.1.0/CHANGELOG.md

@@ -0,0 +1,15 @@
+# Changelog
+
+## 0.1.0
+
+- Add model-independent `Backend`, `Client`, and `serve` APIs for policy, value,
+  reward, and other inference adapters.
+- Exchange binary MessagePack messages and numeric NumPy arrays over WebSocket
+  using the `inferport.v1` protocol.
+- Define connection ownership, reset and cleanup behavior, bounded network waits,
+  error responses, optional bearer authentication, and TLS support.
+- Support Python 3.10+ and NumPy >=1.21.3,<3 with three runtime dependencies.
+- Include examples, protocol documentation, compatibility tests, and benchmarks.
+
+Real model and robot integrations remain in their owning repositories. See
+[validation](docs/validation.md) for the tested scope and remaining gaps.

inferport-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 InferPort contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

inferport-0.1.0/PKG-INFO

@@ -0,0 +1,190 @@
+Metadata-Version: 2.5
+Name: inferport
+Version: 0.1.0
+Summary: Lightweight inference exchange between model backends and execution environments.
+Project-URL: Homepage, https://github.com/jeremy775885/InferPort
+Project-URL: Documentation, https://github.com/jeremy775885/InferPort/blob/main/README.md
+Project-URL: Issues, https://github.com/jeremy775885/InferPort/issues
+Project-URL: Changelog, https://github.com/jeremy775885/InferPort/blob/main/CHANGELOG.md
+License-Expression: MIT
+License-File: LICENSE
+Requires-Python: >=3.10
+Requires-Dist: msgpack<2,>=1.1
+Requires-Dist: numpy<3,>=1.21.3
+Requires-Dist: websockets<18,>=16.1.1
+Description-Content-Type: text/markdown
+
+# InferPort
+
+A small Python library for exchanging inference inputs and results between a model
+repository and a robot, simulator, or evaluation program. Supports policy, value,
+reward, and other backends without importing a model framework or robot SDK.
+
+The public API is `Backend`, `serve`, and `Client`. The execution loop belongs to the
+calling application. Transport is WebSocket with binary MessagePack and NumPy arrays.
+One connection owns one backend; batch is simply part of your array shapes.
+
+## Install and run
+
+Python **3.10+**, NumPy **>=1.21.3,<3**. Runtime dependencies are NumPy, msgpack,
+and websockets. Install a published release from PyPI with:
+
+```bash
+pip install inferport
+```
+
+For installation from source and local development:
+
+```bash
+# From this repository; uv is a development convenience, not a runtime requirement.
+uv sync --group dev
+uv run examples/serve_value.py
+
+# In another terminal, from this repository:
+uv run examples/call_value.py
+# [7. 7. 7. 7.]
+```
+
+Install into either consuming repository's environment with `uv pip install /path/to/InferPort`.
+The repository is named `InferPort`; the installed distribution and import are `inferport`.
+Existing projects can retain a compatible NumPy version, including `1.21.3` on Python 3.10,
+`1.23.5` on Python 3.11, and `1.26.4` on Python 3.10–3.12. The selected NumPy version must
+also support the environment's Python version; newer Python versions need newer NumPy releases.
+
+Model repository:
+
+```python
+import numpy as np
+from inferport import Backend, InvalidInput, serve
+
+
+class ValueBackend(Backend):
+    def infer(self, inputs):
+        state = inputs.get("state")
+        if not isinstance(state, np.ndarray) or state.ndim != 2:
+            raise InvalidInput("expected state with shape [B, D]")
+        return {"value": np.sum(state * state, axis=-1)}
+
+
+serve(ValueBackend())  # localhost:8000; blocks until Ctrl+C
+```
+
+Execution repository:
+
+```python
+import numpy as np
+from inferport import Client
+
+with Client("ws://127.0.0.1:8000") as client:
+    result = client.infer({"state": np.ones((4, 7), dtype=np.float32)})
+    print(result["value"])
+```
+
+For a robot or simulator, call `infer()` inside your own loop and consume the returned
+actions there. Define image layout, joint order, units, action chunk handling, and
+normalization in your adapters. InferPort does not resize images or control hardware.
+
+## State and errors
+
+Only `Backend.infer(inputs) -> dict` is required. Stateful backends implement
+`reset(context) -> None` to clear **all** model/processor/cache state and replace the
+context. An empty context must always work. See [the counter example](https://github.com/jeremy775885/InferPort/blob/main/examples/stateful_counter.py).
+`close() -> None` releases backend resources at service shutdown.
+
+For engines that must be created and used on the same thread, see
+[the thread-bound model example](https://github.com/jeremy775885/InferPort/blob/main/examples/thread_bound_backend.py). It initializes the
+model on the backend worker before READY; later resets retain the loaded model.
+Allow sufficient `open_timeout` for first-time model loading.
+
+The server calls `reset({})` before READY and after disconnect, then `close()` once
+when it exits. All hooks run serially on one worker thread; the Backend constructor
+runs in the caller. `serve()` owns the backend even if startup fails. A second connection
+receives `RemoteError(code="busy")`, including while old work is finishing or cleaning up.
+If initialization or disconnect cleanup fails, the service stops.
+
+`Client()` performs no I/O. Use `with` or explicitly call `connect()` and `close()`.
+Client calls must be sequential; concurrent calls raise `RuntimeError`. Cross-thread
+`close()` interrupts network waits. Closing is idempotent. Create a new Client after
+closing or a fatal failure; there is no reconnect, retry, or automatic request replay.
+
+- Local invalid input types raise `TypeError`/`ValueError` before sending; the connection stays usable.
+- A backend raises `InvalidInput` **before changing state** for an expected input problem.
+  The caller gets nonfatal `RemoteError(code="invalid_input")` and may continue.
+  Its message is public validation guidance (bounded to 512 characters).
+- Unexpected backend exceptions and unsupported outputs produce fatal `RemoteError`.
+  These unexpected errors use generic client messages; server tracebacks are logged locally.
+- `ProtocolError` rejects malformed messages or mismatched responses.
+- `TransportError` covers connection failures; `RequestTimeout` is its subclass and
+  provides `.stage` (`open`, `ready`, `send`, or `response`).
+- All library exceptions inherit `Error`. `RemoteError` exposes `code`, `message`,
+  `request_id`, and `fatal`.
+
+## Deadlines and limits
+
+`Client(uri, timeout=30, open_timeout=10, max_message_bytes=64*1024*1024)`:
+
+- `open_timeout` covers connection, handshake, and READY together.
+- `infer(data, timeout=...)` / `reset(context, timeout=...)` cover send and receive,
+  including model execution. Local encoding and decoding are outside this deadline.
+- Timeout makes the Client unusable. Network cleanup has a separate five-second budget.
+  A timeout cannot cancel an already running backend operation. Its result is dropped;
+  ownership remains held until that work and cleanup finish.
+- A permanently stuck backend requires external process termination. InferPort cannot
+  safely kill GPU/Python worker operations or promise hard realtime behavior.
+
+Payloads are string-keyed dictionaries containing nested dictionaries/lists, basic
+scalars, bytes, and real numeric/bool NumPy arrays. Tuples become lists; NumPy scalars
+become Python scalars. Arrays preserve values, shape, and type width and arrive as
+writable, C-contiguous, native-endian arrays. Object, structured, complex, text,
+datetime, and custom array dtypes are rejected. Convert Torch/JAX tensors explicitly.
+
+The default **64 MiB complete message** limit applies on both sides, including the
+envelope. Configure the same limit on Client and serve (minimum 128 bytes). Arrays
+have at most 32 dimensions; messages at most 32 nesting levels and 100,000 nodes.
+This is not a total process memory limit: serialization, buffers, and writable arrays
+require additional memory. No automatic chunking or compression is performed.
+
+## Deployment
+
+`serve(backend, host="127.0.0.1", port=8000, token=None, ssl=None,
+max_message_bytes=64*1024*1024, send_timeout=30, stop_event=None)`.
+Use `threading.Event` for an embedded service's stop request; Ctrl+C works in a script.
+`send_timeout` bounds network writes, not backend execution.
+
+To listen across machines, explicitly select the listening address. Both sides accept
+`token="..."` for bearer authentication and an `ssl.SSLContext` for TLS. The server
+context must load its certificate; `wss://` clients verify certificates and hostnames
+by default. Use a client context for a private CA. Tokens must be nonempty printable
+ASCII without whitespace; keep them out of URLs. Use TLS or a trusted encrypted tunnel
+across untrusted networks; a token does not encrypt `ws://` traffic.
+
+The only endpoint is `/`, with required subprotocol `inferport.v1`. Browser Origin
+requests are rejected. Compression and system proxy discovery are disabled. Heartbeats
+check connection liveness, not model progress. Standard Python logging provides request
+IDs, operations, durations, and errors without automatically logging payloads.
+
+## Development and scope
+
+```bash
+uv run pytest -q
+uv run ruff check .
+uv run ruff format --check .
+uv build
+uv run benchmarks/roundtrip.py --iterations 100
+```
+
+See [the full protocol and design](https://github.com/jeremy775885/InferPort/blob/main/docs/inferport-design.md),
+[verification results and remaining gaps](https://github.com/jeremy775885/InferPort/blob/main/docs/validation.md), and
+[the maturity assessment and roadmap](https://github.com/jeremy775885/InferPort/blob/main/docs/roadmap.md).
+Maintainers can follow [the release guide](https://github.com/jeremy775885/InferPort/blob/main/docs/releasing.md); changes are recorded in
+[the changelog](https://github.com/jeremy775885/InferPort/blob/main/CHANGELOG.md).
+CI covers Python 3.10–3.14, minimum dependencies (NumPy 1.21.3), NumPy 1.23.5 / 1.26.4,
+and newer combinations.
+It includes Windows/macOS smoke jobs; completed runs and their scope are recorded in
+[the validation record](https://github.com/jeremy775885/InferPort/blob/main/docs/validation.md#github-actions).
+
+InferPort replaces `policy_runtime` with no compatibility alias or TCP/JSON fallback.
+This repository doesn't implement RTC, robot/environment base classes, action scheduling,
+automatic batching, multiple sessions/models, or a schema/description framework.
+Real model and robot integrations remain in their owning repositories and require
+separate validation.

inferport-0.1.0/README.md

@@ -0,0 +1,174 @@
+# InferPort
+
+A small Python library for exchanging inference inputs and results between a model
+repository and a robot, simulator, or evaluation program. Supports policy, value,
+reward, and other backends without importing a model framework or robot SDK.
+
+The public API is `Backend`, `serve`, and `Client`. The execution loop belongs to the
+calling application. Transport is WebSocket with binary MessagePack and NumPy arrays.
+One connection owns one backend; batch is simply part of your array shapes.
+
+## Install and run
+
+Python **3.10+**, NumPy **>=1.21.3,<3**. Runtime dependencies are NumPy, msgpack,
+and websockets. Install a published release from PyPI with:
+
+```bash
+pip install inferport
+```
+
+For installation from source and local development:
+
+```bash
+# From this repository; uv is a development convenience, not a runtime requirement.
+uv sync --group dev
+uv run examples/serve_value.py
+
+# In another terminal, from this repository:
+uv run examples/call_value.py
+# [7. 7. 7. 7.]
+```
+
+Install into either consuming repository's environment with `uv pip install /path/to/InferPort`.
+The repository is named `InferPort`; the installed distribution and import are `inferport`.
+Existing projects can retain a compatible NumPy version, including `1.21.3` on Python 3.10,
+`1.23.5` on Python 3.11, and `1.26.4` on Python 3.10–3.12. The selected NumPy version must
+also support the environment's Python version; newer Python versions need newer NumPy releases.
+
+Model repository:
+
+```python
+import numpy as np
+from inferport import Backend, InvalidInput, serve
+
+
+class ValueBackend(Backend):
+    def infer(self, inputs):
+        state = inputs.get("state")
+        if not isinstance(state, np.ndarray) or state.ndim != 2:
+            raise InvalidInput("expected state with shape [B, D]")
+        return {"value": np.sum(state * state, axis=-1)}
+
+
+serve(ValueBackend())  # localhost:8000; blocks until Ctrl+C
+```
+
+Execution repository:
+
+```python
+import numpy as np
+from inferport import Client
+
+with Client("ws://127.0.0.1:8000") as client:
+    result = client.infer({"state": np.ones((4, 7), dtype=np.float32)})
+    print(result["value"])
+```
+
+For a robot or simulator, call `infer()` inside your own loop and consume the returned
+actions there. Define image layout, joint order, units, action chunk handling, and
+normalization in your adapters. InferPort does not resize images or control hardware.
+
+## State and errors
+
+Only `Backend.infer(inputs) -> dict` is required. Stateful backends implement
+`reset(context) -> None` to clear **all** model/processor/cache state and replace the
+context. An empty context must always work. See [the counter example](https://github.com/jeremy775885/InferPort/blob/main/examples/stateful_counter.py).
+`close() -> None` releases backend resources at service shutdown.
+
+For engines that must be created and used on the same thread, see
+[the thread-bound model example](https://github.com/jeremy775885/InferPort/blob/main/examples/thread_bound_backend.py). It initializes the
+model on the backend worker before READY; later resets retain the loaded model.
+Allow sufficient `open_timeout` for first-time model loading.
+
+The server calls `reset({})` before READY and after disconnect, then `close()` once
+when it exits. All hooks run serially on one worker thread; the Backend constructor
+runs in the caller. `serve()` owns the backend even if startup fails. A second connection
+receives `RemoteError(code="busy")`, including while old work is finishing or cleaning up.
+If initialization or disconnect cleanup fails, the service stops.
+
+`Client()` performs no I/O. Use `with` or explicitly call `connect()` and `close()`.
+Client calls must be sequential; concurrent calls raise `RuntimeError`. Cross-thread
+`close()` interrupts network waits. Closing is idempotent. Create a new Client after
+closing or a fatal failure; there is no reconnect, retry, or automatic request replay.
+
+- Local invalid input types raise `TypeError`/`ValueError` before sending; the connection stays usable.
+- A backend raises `InvalidInput` **before changing state** for an expected input problem.
+  The caller gets nonfatal `RemoteError(code="invalid_input")` and may continue.
+  Its message is public validation guidance (bounded to 512 characters).
+- Unexpected backend exceptions and unsupported outputs produce fatal `RemoteError`.
+  These unexpected errors use generic client messages; server tracebacks are logged locally.
+- `ProtocolError` rejects malformed messages or mismatched responses.
+- `TransportError` covers connection failures; `RequestTimeout` is its subclass and
+  provides `.stage` (`open`, `ready`, `send`, or `response`).
+- All library exceptions inherit `Error`. `RemoteError` exposes `code`, `message`,
+  `request_id`, and `fatal`.
+
+## Deadlines and limits
+
+`Client(uri, timeout=30, open_timeout=10, max_message_bytes=64*1024*1024)`:
+
+- `open_timeout` covers connection, handshake, and READY together.
+- `infer(data, timeout=...)` / `reset(context, timeout=...)` cover send and receive,
+  including model execution. Local encoding and decoding are outside this deadline.
+- Timeout makes the Client unusable. Network cleanup has a separate five-second budget.
+  A timeout cannot cancel an already running backend operation. Its result is dropped;
+  ownership remains held until that work and cleanup finish.
+- A permanently stuck backend requires external process termination. InferPort cannot
+  safely kill GPU/Python worker operations or promise hard realtime behavior.
+
+Payloads are string-keyed dictionaries containing nested dictionaries/lists, basic
+scalars, bytes, and real numeric/bool NumPy arrays. Tuples become lists; NumPy scalars
+become Python scalars. Arrays preserve values, shape, and type width and arrive as
+writable, C-contiguous, native-endian arrays. Object, structured, complex, text,
+datetime, and custom array dtypes are rejected. Convert Torch/JAX tensors explicitly.
+
+The default **64 MiB complete message** limit applies on both sides, including the
+envelope. Configure the same limit on Client and serve (minimum 128 bytes). Arrays
+have at most 32 dimensions; messages at most 32 nesting levels and 100,000 nodes.
+This is not a total process memory limit: serialization, buffers, and writable arrays
+require additional memory. No automatic chunking or compression is performed.
+
+## Deployment
+
+`serve(backend, host="127.0.0.1", port=8000, token=None, ssl=None,
+max_message_bytes=64*1024*1024, send_timeout=30, stop_event=None)`.
+Use `threading.Event` for an embedded service's stop request; Ctrl+C works in a script.
+`send_timeout` bounds network writes, not backend execution.
+
+To listen across machines, explicitly select the listening address. Both sides accept
+`token="..."` for bearer authentication and an `ssl.SSLContext` for TLS. The server
+context must load its certificate; `wss://` clients verify certificates and hostnames
+by default. Use a client context for a private CA. Tokens must be nonempty printable
+ASCII without whitespace; keep them out of URLs. Use TLS or a trusted encrypted tunnel
+across untrusted networks; a token does not encrypt `ws://` traffic.
+
+The only endpoint is `/`, with required subprotocol `inferport.v1`. Browser Origin
+requests are rejected. Compression and system proxy discovery are disabled. Heartbeats
+check connection liveness, not model progress. Standard Python logging provides request
+IDs, operations, durations, and errors without automatically logging payloads.
+
+## Development and scope
+
+```bash
+uv run pytest -q
+uv run ruff check .
+uv run ruff format --check .
+uv build
+uv run benchmarks/roundtrip.py --iterations 100
+```
+
+See [the full protocol and design](https://github.com/jeremy775885/InferPort/blob/main/docs/inferport-design.md),
+[verification results and remaining gaps](https://github.com/jeremy775885/InferPort/blob/main/docs/validation.md), and
+[the maturity assessment and roadmap](https://github.com/jeremy775885/InferPort/blob/main/docs/roadmap.md).
+Maintainers can follow [the release guide](https://github.com/jeremy775885/InferPort/blob/main/docs/releasing.md); changes are recorded in
+[the changelog](https://github.com/jeremy775885/InferPort/blob/main/CHANGELOG.md).
+CI covers Python 3.10–3.14, minimum dependencies (NumPy 1.21.3), NumPy 1.23.5 / 1.26.4,
+and newer combinations.
+It includes Windows/macOS smoke jobs; completed runs and their scope are recorded in
+[the validation record](https://github.com/jeremy775885/InferPort/blob/main/docs/validation.md#github-actions).
+
+InferPort replaces `policy_runtime` with no compatibility alias or TCP/JSON fallback.
+This repository doesn't implement RTC, robot/environment base classes, action scheduling,
+automatic batching, multiple sessions/models, or a schema/description framework.
+Real model and robot integrations remain in their owning repositories and require
+separate validation.