Keep voice tools behind authenticated workspace approvals
Separate caller identity, conversational intent, private tool access and exact approval in a voice-capable agent.
Separate recordings, signed transcripts, operational timings and workspace content policy when investigating calls.
A voice system can produce raw recordings, live audio frames, partial transcripts, final transcripts, tool outputs and timing events. Those items have different sensitivity and usefulness. Start with the narrowest evidence that can answer the question. A carrier connection failure may need only a phase and provider status; it rarely requires the words the caller spoke. Record an explicit purpose for each retained content type and identify who can access it. Avoid copying entire call objects into unrestricted logs just because that makes an early debugging session convenient.
Audio recording is off by default in the documented voice flow. That does not imply there is no transcript or no provider processing. Signed ElevenLabs post-call transcripts can appear in call details and number activity, subject to the workspace content retention policy. The cascade’s recognition URL disables provider logging, while other transport and provider configurations require their own review. State these boundaries accurately when informing users. A claim that no conversation data exists anywhere is broader than the checked implementation and obscures the actual controls an operator needs to inspect.
Stack & timings can expose configured components, provider-reported measurements, backend timing scopes and failure phases without reproducing the conversation. Prefer those fields for latency and transport investigations. Keep private identifiers in the access-controlled incident record rather than a public issue. If a transcript is necessary to understand a recognition error, extract a synthetic reproduction or the smallest authorized excerpt. Do not include provider keys, SIP passwords, personal phone numbers or private voice identifiers. Redaction should happen before the report leaves the workspace’s controlled environment.
Post-call data can arrive after the conversation ended and after a retention decision. Lupin checks the content storage policy before writing signed transcript content. An expired resource should not be repopulated merely because a provider retries an old callback. Include delayed and duplicate delivery in retention tests, alongside a normal timely transcript. Keep the fact that content expired distinct from an invalid signature or storage outage, because these conditions require different operator responses. Retention behavior must remain meaningful even when delivery order and provider timing are imperfect.
A useful private packet contains the workspace resource reference, time range, captured transport, narrow failure or timing fields and the expected behavior. Add synthetic dialogue text where possible and specify when temporary content should be deleted. Restrict access to those resolving the issue and check provider retention settings through the authorized operator. For legal consent requirements, obtain guidance appropriate to the operating region rather than treating this engineering guide as a universal consent rule. The practical objective is enough evidence to fix the defect with a traceable lifecycle for any conversation content retained.
If an incident involves a misheard appointment date, reproduce it with an invented date and neutral sentence before sharing evidence outside the workspace. Preserve only the timing or recognition pattern needed to diagnose the defect. The synthetic example can support regression testing without making a customer’s original conversation part of permanent project history.
Reviewed against current checked-in Lupin routes, shared contracts and official provider references. No live calls, OAuth authorizations, app actions or purchases were performed.
Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.