Live API: fractional-second Durations ("7.280s") fail to deserialize, dropping the whole message

[davetraction]

Description

CustomDurationDeserializer in JsonSerializable only parses whole-second Duration values (Long.parseLong), but the Gemini Live API server streams fractional-second Durations (e.g. "7.280s", "1.600s", "0.360s"). The parse throws NumberFormatException, which surfaces as:

SEVERE: Error deserializing message
com.google.genai.errors.GenAiIOException: Failed to deserialize the JSON node.
Caused by: com.fasterxml.jackson.databind.exc.InvalidFormatException:
  Cannot deserialize value of type `java.time.Duration` from String "7.280s":
  Cannot parse duration from string: 7.280s. Expected format 'Xs'.

Because this happens inside the Live WebSocket read path, the entire server message is dropped. When that message carries the model turn / a function call, the turn is lost — in a voice-agent app the assistant cuts off mid-sentence and no tool call fires.

Impact

Production voice-reception app on gemini-3.1-flash-live-preview (Vertex): ~700 of these errors in 24h across ~190 calls (~3–4 per call). Most calls survive because the read loop resumes on the next message, but an intermittent few break at the worst moment (mid first response / on the transfer tool call).

Root cause

src/main/java/com/google/genai/JsonSerializable.java, CustomDurationDeserializer:

long seconds = Long.parseLong(secondsPart);   // rejects "7.280"
return java.time.Duration.ofSeconds(seconds);

Suggested fix

Parse decimal seconds and keep sub-second precision:

double seconds = Double.parseDouble(secondsPart);
return java.time.Duration.ofNanos(Math.round(seconds * 1_000_000_000.0));

(Protobuf google.protobuf.Duration is routinely JSON-encoded with fractional seconds, so this is the expected wire format.)

Versions

• google-genai 1.61.0 (also present in 1.62.0 — JsonSerializable unchanged)
• Model: gemini-3.1-flash-live-preview, Vertex AI, Live API (BidiGenerateContent)
• Java 25

[hemasekhar-p]

Hi @davetraction, Thank you for bringing this to our attention. We truly appreciate you flagging this issue and your contribution. we will be filing this issue and PR internally for review.