# AI Operations Brief — runbook and live evidence

Answers, using a real MCP client through the real ITWorx MCP Hub connector against live
production Fleet Ops: *"Which vehicles need the most attention today, why, and which
internal procedure should be followed for the most important issue?"*

## What this is not

Not a chatbot. No write actions exist under `/api/v1/integrations/mcp/*` (verified by
`test_no_write_endpoints_exist_under_mcp_namespace`). Every call below is a plain
read-only tool invocation, exactly as the deployed Hub connector performs them.

## Reproducible command

Run from inside the live `itworx-mcp-hub-connector-mobilityops-1` container (has the
real `MOBILITYOPS_ENDPOINT`, `MOBILITYOPS_SERVICE_TOKEN_FILE` and the real
`packages.connector_kit.mobilityops.MobilityOpsClient` already available — the same
client class the deployed connector uses):

```sh
docker exec -e PYTHONPATH=/opt/hub -w /opt/hub \
  itworx-mcp-hub-connector-mobilityops-1 python3 - <<'PY'
from packages.connector_kit.mobilityops import FileTokenProvider, MobilityOpsClient, MobilityOpsSettings
import os, json

settings = MobilityOpsSettings(
    base_url=os.environ["MOBILITYOPS_ENDPOINT"],
    token_reference=os.environ["MOBILITYOPS_SERVICE_TOKEN_FILE"],
)
client = MobilityOpsClient(settings, FileTokenProvider())
client_id = "ai-ops-brief"

summary = client.operations_summary(client_id)
attention = client.attention_vehicles(client_id, minimum_severity="high", date=None, limit=20)
top_ref = attention[0]["vehicle_ref"] if attention else None
details = client.vehicle_details(client_id, top_ref) if top_ref else None
answer = client.search_knowledge(client_id, "What must I do when a vehicle returns with damage?", max_sources=2)

print(json.dumps({"summary": summary, "top_attention": attention[0] if attention else None, "vehicle": details, "answer": answer}, indent=2))
client.close()
PY
```

This calls, in order: `fleet_ops_get_operations_summary` → `fleet_ops_list_attention_vehicles`
→ `fleet_ops_get_vehicle_details` → `fleet_ops_search_knowledge`. No result is invented —
every field printed is exactly what Fleet Ops's live API returned.

## Live run, 2026-08-05

- **Operations summary**: 21 available, 11 rented, 6 cleaning, 5 maintenance,
  **7 blocked**; 23 open quality issues; 1 pending/failed workflow.
- **Most pressing vehicle**: `MO-031` — `missing_required_field`, "near-future booking;
  required operational inspection missing", detected `2026-08-05T11:08:33Z`. (16 vehicles
  currently carry a `high`-severity open issue; `MO-031` was the earliest-detected.)
- **Vehicle detail (`MO-031`)**: Adria Matrix, 2022, Geel, `operational_status: blocked`,
  41,149 km (service due at 50,000 km), no current booking.
- **Grounded procedure (English)**: *Damage handling procedure* v1.3, section "1.
  Immediate actions" — "When a vehicle returns with visible or reported damage, mark
  damage in the return inspection, add a concise factual description and keep the
  vehicle blocked. Do not promise the customer a repair cost or liability decision."
  Second source: *Vehicle return procedure* v2.0, section "3. Determine next state".
  `evidence_state: grounded`, 2 real citations, `provider: demo`.
- **Dutch and French variants of the damage question** (`Wat moet ik doen wanneer een
  voertuig beschadigd terugkomt?`, `Que dois-je faire lorsqu'un véhicule revient
  endommagé ?`) both returned `evidence_state: insufficient` — an honest, non-fabricated
  "no match" rather than a wrong or invented answer. Root cause: the currently-deployed
  Hub connector does not yet send the new `locale` field this session added to
  `search-knowledge` (Fleet Ops defaults to `en-GB`), so non-English question text
  doesn't match the demo provider's English-tokenized index. A real fix needs a Hub-side
  connector update to pass `locale`, tracked as a follow-up, not silently worked around.
- **Correlation ID, end to end, verified**: each call's response `correlation_id` (e.g.
  `abd643a7-2aeb-4af3-806a-da04acb3e444` for the English grounded answer) appears
  verbatim in Fleet Ops's own audit log (`GET /api/v1/audit?action=mcp_tool_request`),
  alongside `actor_label: ai-ops-brief-2026-08-05` and the real tool name
  (`fleet_ops_search_knowledge`). No write actions were performed; only `mcp_tool_request`
  audit rows were created, matching every other live MCP call this integration makes.

## Known limitation, stated plainly

The demo knowledge provider (not RAGcore — `KNOWLEDGE_PROVIDER=demo`, see
`docs/final-integrations/current-state-audit.md` for why) is what grounds the English
answer here. It is deterministic, extractive, and never fabricates — but it is not the
live RAGcore integration the brief brief for this task set out to exercise; that
remains blocked on RAGcore's own reranker gap. This run is honest about that: it proves
the full MCP-client → Hub → Fleet Ops → knowledge-provider → audit chain works for real,
live, in production, with real data and real citations — using the knowledge provider
that is actually configured live today.