Resume WebSocket live testing

The WebSocket transport and native CES connector are implemented. The October 17, 2026 development rollout runs runtime commit cda96b7. A real WxCC call completed two caller turns over native CES WebSocket, with both turns confirmed by provider recognition events. A retained gRPC control also completed two recognized turns. Full live acceptance still requires the remaining matrix checks below. This gateway remains functional sample code.

The live run exposed a streamed-prompt boundary defect: provider audio CHUNKs could finish with a prompt-less FINAL, leaving WxCC in playback and dropping caller input. The sender now terminates that stream with explicit empty audio, including an abandoned stream before caller speech starts. Its synthetic FINAL preserves the previous audio frame’s input collection settings. Local regression coverage includes normal and terminal responses, including interleaved input-boundary events. The deployed image passed its registry scan with no reported findings; gateway tests passed (644 tests), and caller-tool tests passed (118 tests).

Observed development results (October 17–18, 2026)

The private development profile permits interruption of autonomous CES reprompts and uses an agent-specific caller prompt gate. Greetings and caller-triggered replies retain their existing non-interruptible behavior. The former 200 ms recognition-assisted override cut off the second half of the pause fixture; the target now uses the reference configuration’s normal pause window. Recheck latency when changing endpointing settings.

Reliable live DTMF delivery remains an acceptance blocker. Failed calls contain no gateway DTMF input, while successful calls prove the JSON/protobuf conversion and Local Audio handlers work. Hold digit, media and duration fixed while varying submission before/after speech endpointing; the observed correlation does not establish a root cause. Flow Designer’s handoff variable value, active-call shutdown/drain, and load behavior still need their own live checks. Keep exact call artifacts, configuration fingerprints, and deployment identities in the private handoff rather than this shared repository.

Lex WebSocket adapter: local validation only

AWSLexWebSocketConnector is a separate WxCC-facing adapter with distinct bot discovery IDs. It shares the existing utterance-buffered Lex provider and emits complete WAV FINAL responses. Local tests cover discovery and transport isolation, voice, DTMF mapping, SDK request validation and disconnect cleanup. The AWS SDK baseline is Boto3/Botocore 1.43.114 and S3Transfer 0.19.2. This adapter has not been deployed or validated through a real WxCC-to-Lex call. See Lex setup.

1. Establish the local baseline

From the repository root, activate a virtual environment with requirements-dev.txt installed, then run:

source venv/bin/activate
python -m grpc_tools.protoc -I./proto \
  --python_out=src/generated --grpc_python_out=src/generated proto/*.proto
python -m pytest tests/

The SDK caller and its tests are maintained separately in the private webex-test-caller application (repository access required). Follow its setup guide and run its tests from that checkout, using its own virtual environment.

For a local socket smoke check, start the gateway in one terminal:

GATEWAY_CONFIG=config/websocket_local_example.yaml python main.py

In another terminal with the same virtual environment activated:

python tools/websocket_smoke.py --url ws://127.0.0.1:58765 --mode list
python tools/websocket_smoke.py --url ws://127.0.0.1:58765 --mode session \
  --agent-id 'Local Audio: Local Playback'

Discovery must list Local Audio. The session probe now reads through FINAL, checks conversation/sequence ordering, validates base64 audio and requires the explicit empty-audio terminator after streamed audio. Early disconnects, error frames and incomplete streams fail the command. --timeout bounds the whole response, so a trickle of frames cannot keep an incomplete probe alive forever. Default session output contains only conversation IDs and frame/audio byte counts; --include-audio explicitly prints complete frames with call content.

Use a small concurrent run against the local listener to check that sessions complete and remain open for five seconds:

python tools/websocket_smoke.py --url ws://127.0.0.1:58765 --mode session \
  --agent-id 'Local Audio: Local Playback' --sessions 8 --hold-seconds 5

For a shutdown check, replace the hold setting with --hold-seconds 30 --expect-shutdown, then stop the local gateway with Control-C after eight session_ready progress events appear on stderr. Every session must finish its initial response and close with 1001 within the hold window. A normal close, early disconnect or missing shutdown fails. Check gateway/provider cleanup separately: a 1001 receipt alone does not prove that their resources were released. --sessions is limited to 1–100 and hold time to 0–300 seconds; each concurrent call gets a unique ID. BYOVA_WEBSOCKET_TOKEN supplies the existing authentication header for an approved authenticated target without placing the token on the command line.

These direct socket checks do not place a phone call or certify WxCC media, DTMF, provider recognition, Flow variables or production capacity. The local socket regression tests also check shutdown with eight active conversations, one provider cleanup per conversation and no remaining registry leases.

2. Prepare the live target when returning

• Record the selected source commit and verify the target runs that same commit. Build releases only with scripts/build-runtime-release.sh --ref HEAD --output /tmp/byova-gateway-runtime.tar.gz after committing the intended changes.
• Use the existing private deployment procedure and configuration. Keep hostnames, resource IDs, credentials, and routing outside this shared example repository.
• Verify WSS routing for both /v1/va and /v1/listVirtualAgents. Enable enforced JWT validation and the distinct WebSocket datasource/Service App profile; the loopback-only unauthenticated example is unsuitable for the live target.
• Select the intended provider hop explicitly. GECXConnector uses CES gRPC; GECXWebSocketConnector uses CES WebSocket. Give them distinct agent IDs and transport restrictions when configured together. See the GECX setup guide.
• Use a dedicated test caller and approved WxCC test flow/destination. Configure the E2E runner’s private gateway-events access as described in its JSON test and gateway-evidence guide.

3. Run the acceptance matrix

The existing E2E plans already cover the Local Audio terminal controls and GECX voice/terminal scenarios. From the separate webex-test-caller checkout, using its configured environment:

webex-test-caller validate --config config/local-audio-dtmf.spec.json --list
webex-test-caller validate --config config/gecx-regression.spec.json --list

Run each Local Audio test against the test destination:

webex-test-caller run --destination "$BYOVA_TEST_DESTINATION" \
  --config config/local-audio-dtmf.spec.json --test dtmf-transfer
webex-test-caller run --destination "$BYOVA_TEST_DESTINATION" \
  --config config/local-audio-dtmf.spec.json --test dtmf-session-end

Run the applicable listed GECX test IDs with the same --config/--test pattern, adding --gateway-transport websocket --gateway-agent-id "$BYOVA_TEST_AGENT_ID" to bind evidence to the selected native CES WebSocket agent. Supply these values locally; do not commit deployment-specific test plans. Provider prompts and handoff expectations must match the selected agent’s actual behavior.

4. Record acceptance

For every run, retain the deployed commit, non-secret configuration fingerprint, test ID, UTC time window, conversation ID, gateway transport, provider agent ID, caller result, gateway terminal/cleanup events, and provider session evidence. Keep recordings, transcripts, raw logs, and private deployment details in the approved private evidence location.

A caller recording alone does not prove provider cleanup or handoff metadata. For every speech scenario, confirm the expected number of provider recognition events as well as caller audio. An autonomous no-input reprompt can satisfy an audio-only response expectation even when the caller utterance was discarded. Tune the private profile’s remote-silence gate for the selected agent. Short intervals can inject within a greeting; long intervals can overlap a no-input reprompt. Record that interval with the configuration fingerprint. Mark each matrix row pass/fail/pending and link its evidence. A provider cleanup that exceeds the drain budget is a failed shutdown check; Python cannot forcibly stop a blocked provider thread. Do not label the feature live-verified until every applicable row passes on the tested commit.

The feature’s Spec Kit-compatible acceptance criteria and task status are under specs/001-websocket-transport.