Summary

Closes #11. When loading a model via Ollama there was no signal whether execution runs on GPU or CPU. This adds a runtime status object that reports the execution device, and surfaces it from the load CLI.

Changes

• Domain (domain/model.py): added Device enum (gpu / cpu / unknown) and a pure RuntimeStatus pydantic model (backend, available, device, loaded_models, details). No I/O, per the architecture's purity rule.
• Port (ports/runtime.py): added status() -> RuntimeStatus to RuntimePort.
• Base runtime (adapters/runtimes/base.py): status() reports availability + device; _detect_device() hook defaults to UNKNOWN so other runtimes get the contract for free.
• Ollama adapter (adapters/runtimes/ollama.py): _detect_device() calls client.ps() and inspects size_vram per loaded model — > 0 => GPU, 0/None => CPU, no models loaded => UNKNOWN. Best-effort: a failing ps() never breaks the status probe.
• Manager + CLI (core/manager.py, cli/commands/load.py): ModelManager.runtime_status() delegates to the runtime; load now prints Device: <gpu|cpu|unknown> after a successful load.

Acceptance criteria (from #11)

• Add a device/executor field to the runtime status object — RuntimeStatus.device
• Ollama adapter populates it from ollama ps / API metadata — OllamaRuntime._detect_device()
• CLI load prints the device — cli/commands/load.py
• Tests with a fake port — unit + port-contract + integration

Tests

• Unit (tests/unit/test_ollama_runtime.py): GPU (vram>0), CPU (vram=0), UNKNOWN (no models), UNKNOWN (ps raises), unavailable (daemon down) — all with a fake client.
• Port-contract (tests/unit/test_port_contract.py): status() returns a valid RuntimeStatus with a known device value; FakeRuntime in conftest.py gained status().
• Integration (tests/integration/test_ollama.py): test_ollama_status_device_detection loads a local model and cross-checks the detected device against ollama ps (auto-skips without a daemon; cloud/subscription models are filtered out).

Verification

Ran against a live Ollama daemon (ollama serve):

• test_ollama_status_device_detection — passed; correctly reported device=cpu (machine has no usable GPU; ollama ps showed 100% CPU, size_vram: 0).
• Unit suite (ollama + port-contract + domain): 43 passed.
• ruff check clean, mypy --strict clean, ruff format clean.

Notes

• No core/CLI/API contract changes beyond the additive status() method; existing behavior is unchanged.
• The device field is best-effort: it reflects currently-loaded models. If nothing is resident, it reports unknown rather than guessing.

🤖 Generated with Claude