fix(ollama): probe the requested model directly - #128

Summary

An Ollama agent can be reported ready when only a different model is installed. For example, the list entry codellama:7b-instruct satisfies the current substring check for codellama:7b; model names and namespaces can collide the same way.

Use ollama show <model> and its exit status to check the requested model. Ollama resolves omitted tags itself, so codellama still works for an installed codellama:latest. A missing model now gives model_not_found instead of a false ready result.

Scope

• Adapters

Validation

• python3 -m unittest discover -s tests -p 'test_*.py': 471 tests, 4 platform skips
• ./scripts/run_runtime_gate_tests.sh: PASS
• npm test: 19 passed, 1 platform skip

The new regression fails on main for three model-name collisions and passes with this change. A local executable fixture also exercises mco agent check local-coder --json and confirms ready: false for an unavailable model.

Artifacts

• Gate report: reports/adapter-contract/2026-10-04/runtime-gh-report.md (generated locally)

Risk and Rollback

Low risk: the command changes only the model-presence probe, and does not pull or run a model. Revert this commit to restore the previous probe.

All checks above ran locally with Python 3.14.