Reviewing Real Conversations

Load poly-adk-workflow first for the build loop this feeds back into. Once an agent handles real calls, the question becomes "what happened on that call?" — and answering it depends on instrumentation added while building. Logs and metrics are not available retroactively.

Inspecting conversations

poly conversations list                        # recent conversations: ID, start, duration, caller, channel, variant, handoff, summary
poly conversations list --limit 20 --offset 10
poly conversations get <conversation_id>       # full metadata + turn-by-turn transcript
poly conversations get-audio <conversation_id> -o recording.wav

• get includes channel, language, duration, handoff, tags, PolyScore, summary, and every turn — use --json when analyzing programmatically.
• get-audio downloads the WAV for a voice call; --direction user|agent|combined picks the track, --redacted fetches the redacted version.
• The CLI reads individual conversations. Aggregate views — containment, CSAT, handle time, flagged transcripts — live in Agent Studio, not the CLI.

Treat what you find as input to development: a surprising real conversation is the raw material for the next test_suite/ case (see poly-adk-testing). What callers actually say is reliably different from what you imagined.