feat(regen): listen V1 reconfiguration and agent custom think-provider messages

[GregHolmes]

Summary

• Regenerate the Java SDK with fernapi/fern-java-sdk 4.23.1.
• Preserve all 23 required compatibility patches and restore their .fernignore entries.
• Retain generated Listen V1 Configure/Error and Agent custom Think-provider surfaces.
• Add expanded websocket regression coverage and a runnable Listen V1 reconfiguration example.

Validation

• ./gradlew spotlessCheck test compileExamples
• Tier 1 live integration tests passed for Read, Management, Speak REST, and Listen transcription.
• Temporary live smoke with DEEPGRAM_API_KEY: Listen V1 Configure rejection dispatched as typed ListenV1Error; Agent WebSocket authenticated and received Welcome.

[dg-coreylweathers]

Requesting changes for one fix; the rest is small.

What this PR does: Regenerates the Java SDK with Fern 4.23.1 and keeps all 23 hand patches. Live transcription (Listen V1) gets sendConfigure(...), which changes keyterms and feature toggles on a running stream without reconnecting, plus a typed onErrorMessage(...) handler for server Error messages. The Voice Agent socket gets custom Think-provider messages in both directions.

Blocking: server errors on Listen V1 now disappear for anyone who registered only onError.
Expected: a server Error message always reaches a handler the developer registered; on main it arrived through onError.
Observed: with no onErrorMessage registered, V1WebSocketClient.java:564-577 parses the message and drops it. The README quickstart and LiveStreaming.java register only onError. I confirmed against production that a malformed Configure returns exactly this kind of Error (SchemaError).
Fix: when errorHandler == null, pass the error to onErrorHandler; add a test for the only-onError case; record the patch in the AGENTS.md entry for this file.

Should-fix

• ClientOptions.close() Javadoc (lines 147-150) went back to Fern's wording, "any request issued after this method returns fails with a RejectedExecutionException". That's not true for synchronous calls, because close() only shuts down the dispatcher executor. Please restore main's wording and add it to the AGENTS.md ClientOptions entry.
• Version literals are bumped to 0.11.1 in ClientOptions.java and .fern/metadata.json, while pom.xml, build.gradle and the release-please manifest say 0.11.0. Please revert to 0.11.0 and let release-please bump all five files together.

Nits

• .fern/metadata.json: restore the trailing newline.
• LiveReconfigure.java:47,73: print error.getMessage() / e.getMessage() so a failed run shows why.
• README: add a short Listen V1 sendConfigure snippet that also registers onErrorMessage.

What I checked

• Build: spotlessCheck test compileExamples passed in Docker; CI green on Java 11, 17 and 21.
• Hand patches: all 23 present, no .bak files left; the query-param patch is intact in all four streaming clients.
• Live API: production accepts the example's Configure payload (results, then a normal close). The live Error message matches ListenV1Error; code is absent, and it's optional in the type.
• Compatibility: nothing removed, renamed or made required.

Question: the server's error text lists Sync as an accepted client message (CloseStream, Configure, Sync, KeepAlive, Finalize), but it isn't in the AsyncAPI spec or the SDK. Is it public?

[GregHolmes]

Addressed Corey's requested changes in 02cf2e8.

• Listen V1 server Error frames now fall back to onError when no onErrorMessage handler is registered, with regression coverage for that path.
• Restored the accurate ClientOptions.close() Javadoc and 0.11.0 version consistency.
• Added the Listen V1 Configure/onErrorMessage README snippet, improved LiveReconfigure diagnostics, and documented both manual patches in AGENTS.md.
• Restored the metadata trailing newline.

Validation: ./gradlew spotlessCheck test compileExamples.

On Sync: it is absent from the current AsyncAPI and docs, so I left it unmodeled rather than publishing an unverified client message contract. It needs server-team confirmation before adding it to the SDK.

[dg-coreylweathers]

Approving. Commit 02cf2e8 addresses every item from the 2026-10-07 review, and I pushed one more commit on top for the keyterm caveat and the generic-handler exception.

What this PR does. Regenerates the Java SDK with Fern 4.23.1. Listen V1 streaming gains sendConfigure(...) for mid-stream Nova-3 keyterm and numerals changes plus a typed onErrorMessage(...) handler, and the Voice Agent socket gains custom Think-provider messages in both directions. All 23 hand patches survive.

What I checked on 02cf2e8.

• Every earlier finding addressed? Yes. B1: a parsed server Error reaches onError when no onErrorMessage handler is registered, with dispatchesServerErrorsToGenericHandlerWithoutTypedHandler covering that path and AGENTS.md recording the patch. S1: the close() Javadoc matches main again and the ClientOptions entry in AGENTS.md names it. S2: 0.11.0 in ClientOptions.java, pom.xml, build.gradle, .fern/metadata.json, and the release-please manifest. N1 to N3: trailing newline, getMessage() in the example, and the README Configure snippet with onErrorMessage.
• Gates on the new head? Docker eclipse-temurin:17-jdk: ./gradlew spotlessCheck test compileExamples exit 0, 159 tests, 0 failures. CI green on Java 11, 17, and 21.
• Sync: not modeled, and it appears in no spec or public page. Agreed that it waits for server-team confirmation.

What I added in the follow-up commit.

• README.md (above sendConfigure) and the LiveReconfigure.java header now say: keep the keyterm list under the 500-token limit, because an over-limit update currently stops transcription without an Error and the server closes the stream. Verified on production via deepgram-python-sdk #799 (request IDs 01a11601-91c7-7f50-bb58-8f9792fb5b30 and 01a11602-5735-7ce2-a7a2-d34d364a7d15; Linear API-173 still open). The generated ListenV1Configure Javadoc still says the opposite and is not frozen, so it waits for the deepgram-docs spec correction.
• New hand-written ListenV1ErrorException (extends DeepgramApiException) carrying the typed ListenV1Error; the Listen V1 fallback now hands this to onError instead of a bare RuntimeException, so generic-handler callers can read variant, description, and code. The message stays "variant: description", with " (code)" appended when the server sends one. Registered as permanently frozen in .fernignore and AGENTS.md. The regression test now asserts the exception type and the carried fields.
• Gate after the change, Docker eclipse-temurin:17-jdk: spotlessApply made no changes; spotlessCheck test compileExamples exit 0, 159 tests, 0 failures.

Still open elsewhere. The deepgram-docs AsyncAPI description carries the same over-limit sentence until API-173 ships or a docs correction lands.

[dg-coreylweathers]

Approving. Commit 02cf2e8 addresses every item from the 2026-10-07 review, and I pushed one more commit on top for the keyterm caveat and the generic-handler exception.

What this PR does. Regenerates the Java SDK with Fern 4.23.1. Listen V1 streaming gains sendConfigure(...) for mid-stream Nova-3 keyterm and numerals changes plus a typed onErrorMessage(...) handler, and the Voice Agent socket gains custom Think-provider messages in both directions. All 23 hand patches survive.

What I checked on 02cf2e8.

• Every earlier finding addressed? Yes. B1: a parsed server Error reaches onError when no onErrorMessage handler is registered, with dispatchesServerErrorsToGenericHandlerWithoutTypedHandler covering that path and AGENTS.md recording the patch. S1: the close() Javadoc matches main again and the ClientOptions entry in AGENTS.md names it. S2: 0.11.0 in ClientOptions.java, pom.xml, build.gradle, .fern/metadata.json, and the release-please manifest. N1 to N3: trailing newline, getMessage() in the example, and the README Configure snippet with onErrorMessage.
• Gates on the new head? Docker eclipse-temurin:17-jdk: ./gradlew spotlessCheck test compileExamples exit 0, 159 tests, 0 failures. CI green on Java 11, 17, and 21.
• Sync: not modeled, and it appears in no spec or public page. Agreed that it waits for server-team confirmation.

What I added in the follow-up commit.

• README.md (above sendConfigure) and the LiveReconfigure.java header now say: keep the keyterm list under the 500-token limit, because an over-limit update currently stops transcription without an Error and the server closes the stream. Verified on production via deepgram-python-sdk #799 (request IDs 01a11601-91c7-7f50-bb58-8f9792fb5b30 and 01a11602-5735-7ce2-a7a2-d34d364a7d15; Linear API-173 still open). The generated ListenV1Configure Javadoc still says the opposite and is not frozen, so it waits for the deepgram-docs spec correction.
• New hand-written ListenV1ErrorException (extends DeepgramApiException) carrying the typed ListenV1Error; the Listen V1 fallback now hands this to onError instead of a bare RuntimeException, so generic-handler callers can read variant, description, and code. The message stays "variant: description", with " (code)" appended when the server sends one. Registered as permanently frozen in .fernignore and AGENTS.md. The regression test now asserts the exception type and the carried fields.
• Gate after the change, Docker eclipse-temurin:17-jdk: spotlessApply made no changes; spotlessCheck test compileExamples exit 0, 159 tests, 0 failures.

Still open elsewhere. The deepgram-docs AsyncAPI description carries the same over-limit sentence until API-173 ships or a docs correction lands.