The real prompt actually sent to the model on this tenant's most recent turn — not a reconstruction. Read-only; editing the live prompt is a separate, bigger conversation-core question (flag if wanted, not built here).

Pick another tenant to see its config and system prompt side-by-side with this one.

What's DB-driven per tenant/domain today vs. still hardcoded in the pipeline code, and config keys that exist but nothing reads. Static analysis — re-run whenever the pipeline changes (see docs/tenant_config_discovery_checklist_aug20.md).

Tolerant lookup: tries an exact phone match first, then a substring match if nothing matches exactly (mirrors the local trace-viewer tool). "Reset" clears the search view (phone field, results) AND, for a number matching the test-number pattern (9999XXXXXXXXX, the regression harness's own convention), also permanently deletes everything for that number — conversation_traces, messages, states, slots, locks — after a confirmation prompt (fixed Aug 23; it used to only clear the view, never the underlying data, which confused repeated test runs). For any other number, Reset stays view-only, same as before — the backend independently refuses a real delete for a non-test number regardless of what this page does, so this can't touch real customer data. "Delete Test Data" does the identical delete explicitly, kept as its own button for a one-click delete without needing to also reset the view. No replay, no scenario editing (see docs/trace_viewer_ops_panel_aug19.md for why those were left out of this pass). This is a manual/one-shot lookup — turn on Auto-refresh to watch a live conversation advance turn-by-turn without re-clicking Look up.