[fix] - harness console follows the Ollama→LM Studio fallback like /query and /health

[cgfixit]

Proposed changes

Phase 1's primary compat finding — the one real gap in the graceful LM Studio fallback story. The 8-minute scan verified the hot paths are already provider-generic (llm/client.py posts only OpenAI-body /chat/completions, utils/health.py probes generic GET /models, mcp_hybrid_server.py has zero LLM dependence, graph.py's offline_best_effort/local_llm nodes go through the same client) and that resolve_local_backend() + the config.yaml fallback block already implement primary→fallback failover for gate/graph/health. Except the harness console: create_app read models.local_llm.{base_url,model} directly, so with fallback.enabled: true and Ollama down, /query and /health switched to LM Studio while the console kept targeting the dead :11434.

• harness/server.py: route the console through the same llm.client.resolve_local_backend() (new _resolve_backend shim: no network probe when fallback is disabled — the shipped default — and empty/unreadable config still degrades to the Ollama default instead of failing app build). /api/status now reports the resolved provider/base_url/model instead of the raw config primary.
• harness/ollama.py: unreachable-error text was hardcoded "is Ollama running?" — now provider-neutral (details already carry the actual base_url).
• Regression test: fallback enabled + failing primary probe → /api/status must report the lmstudio backend. Proven to fail on the old code (revert-run) and pass on the fix.

Invariant / Governance Impact: None. harness/ (out-of-band) now imports llm.client — a shared library module like utils/, not one of the core three; I6 forbids imports between gate.py/graph.py/mcp_hybrid_server.py and the out-of-band packages, which is untouched. invariant-guard: 28 passed / 0 failed. No graph edge, gate, sanitizer, or soul path changed.

Types of changes

• Bugfix (non-breaking change which fixes an issue)

Scope note: Out-of-band harness layer (Ollama/LM Studio compat).

Benefits / why

• Completes the operator's fallback contract: with fallback.enabled: true, an LM-Studio-only (or Ollama-down) host now gets a working console, matching /query//health behavior — the pre-Ollama-migration workflow keeps working as a backup.
• /api/status stops lying about which backend chat actually talks to.
• Zero behavior change for the shipped default (fallback.enabled: false → primary returned with no probe; all 26 pre-existing harness tests pass unchanged).

Risks to monitor

• The resolver's process-lifetime cache (documented in llm/client.py) means a fallback selection sticks until restart even if the primary recovers — a deliberately unchanged, pre-existing property shared with gate.py; reset_local_backend_cache() exists for tests/reload. Flagged in the scan as an operability note, deliberately not "fixed" here (speculative TTL machinery under feature freeze).
• Note for enablers: with base_url pointed at LM Studio, /health will (correctly) flag a model: value that isn't in the server's /models list — LM Studio ids differ from Ollama tags; set fallback.model per the config comment.

Further comments

Verified: ruff clean; pytest tests/test_harness.py → 27 passed; revert-proof (old server.py fails the new test, fix passes); invariant-guard 28/0. Scan also confirmed three of #415's known migration items are already resolved on main (guardrails config → 11434 with integration.py overriding from GuardrailsConfig; terminal.html message already says Ollama; agentic/config.py provider list already retired lmstudio) — ledger comment going to #415. Companion cleanup PR (orphaned mock_lmstudio.py + a localhost→127.0.0.1 config consistency nit) follows separately.

🤖 Generated with Claude Code

https://claude.ai/code/session_0138hmScu3tFU4v1jzvFUwVT

---

Generated by Claude Code