Skip to content
Chapter 6 · Contributing 6-4

Python source boundary

Status: 2026-09-13 — the repository keeps no Python product source. The last retained provider sources (src/ru_runtime/voice and src/ru_runtime/speech plus the persistence/policy/system mirrors) were removed together with the voice/STT, VTS, and screen-observation media surfaces: the OmniVoice server, the Python benchmarks, the [voice] optional extra, and the mypy src gate no longer exist. Python in this repository is tooling — verification scripts, maintenance helpers, and their tests. Amendment (2026-10-08): Python may also back optional product capabilities (skills declaring requires: python); the default path stays in-process .NET and a missing interpreter fails the capability closed with a named code. Bundling an interpreter requires an ADR.

The deterministic inventory is generated from %RU_ROOT% with:

.venv\Scripts\python.exe scripts\check_python_runtime_boundary.py `
  --root .

The scanner records relative paths, a bounded content hash, byte count, classification, and an is_known_managed_startup label for the two configured C# entrypoint files. Schema 2 reports known_managed_startup_count, not a reachability count. It does not analyze C# launch edges, Python imports, or transitive dependencies. It does not import or execute product modules. The scanner is inventory evidence, not permission to launch a retired module. Output is count-only on stdout. Prompt parity is owned by the C# harness contract tests; this scanner no longer imports a Python prompt implementation or generates a cross-language baseline.

Current boundary classes are:

  • managed-startup: the two configured C# entrypoint files, backend-dotnet/RuBackend/Program.cs and backend-dotnet/RuBackend/Hosting/InProcessExecutionAuthority.cs;
  • tooling: every file under scripts/, tools/, and tests/;
  • legacy: every other scanned file — the rest of backend-dotnet/** plus gui-winui3/**, gui-winui3.Tests/**, and gui-winui3.Harness/** — that is neither an entrypoint nor tooling.

The scan roots are scripts, tools, tests, backend-dotnet, gui-winui3, gui-winui3.Tests, and gui-winui3.Harness.

The interpreter rule is enforced, not just inventoried: the scanner exits 1 when a file with an interpreter suffix (.py, .js, .ts, …) appears under a product root, and CI runs that check directly. The pinned exception is the GUI's vendored browser assets, gui-winui3/Assets/terminal/xterm.js and addon-fit.js: WebView2 parses them for the interactive terminal, they are not a shipped runtime, and InteractiveTerminalContractTests pins their presence. The allowlist is exact paths, so any other .js, .ts, or .py under a product root still fails.

The managed model runtime is the Host's in-process typed model step (ModelStepService behind IModelStepRunner): there is no worker process, loopback port, wire protocol, or provider bootstrap file. Tool execution is in process as well (BrokerExecutionRuntime behind IToolGatewayFactory, ADR-0014), so the Host starts no managed child at all: the Broker child, its loopback lease and RuAgentBroker.exe are gone. There is no device v1 graph, no OmniVoice process, no shared lease coordinator, and no provider credential written to disk. Media adapters (voice, STT, VTS, observation) and their Python counterparts are gone; this inventory is classification evidence, not evidence of OS-level confinement. Terminal children in standard and full_access run as the current Windows user without OS filesystem or network confinement. Structured file tools retain their own path protections.