test-stt-pipeline

name	test-stt-pipeline
description	Run the pipeline-level integration test suite for the STT recognition loop. Tests drive the full run_stt_worker → _recognition_loop → on_stt_event path using a ScriptedBackend (no real microphone or STT model). Use when you want to verify the recognition pipeline is correctly wired — wake detection, command matching, direct triggers, one-breath, and reply window.
license	MIT
metadata	{"author":"[e-mail]","version":"1.0"}

Run the STT pipeline integration tests. These tests exercise the full recognition loop with scripted transcripts — no real microphone, STT model, or network required.

Steps

1. Run the test suite:

   uv run pytest tests/test_pipeline_e2e.py -v
   

2. Report results — for each test class, show pass/fail:

   Scenario	Test	Result
   Two-step wake→command	TestTwoStepWakeCommand	✓/✗
   Wake→no match	TestTwoStepWakeNoMatch	✓/✗
   One-breath	TestOneBreath	✓/✗
   Direct trigger	TestDirectTrigger	✓/✗
   Gated trigger suppressed	TestGatedTriggerNoWake	✓/✗
   Reply window match	TestReplyWindowMatch	✓/✗
   Reply window timeout	TestReplyWindowTimeout	✓/✗

3. On failure — show event trace: When a test fails, the assertion message contains the full event list that was collected. Report it verbatim so it's clear which events fired vs. what was expected. For example:

   FAILED TestTwoStepWakeCommand::test_wake_then_matched
   AssertionError: no matched event; got ['listening', 'level', 'wake', 'level', 'nomatch', 'listening']
   

   This tells you: wake fired correctly, but the transcript didn't match any trigger — likely a fuzzy matching threshold issue or the trigger command string doesn't match the scripted transcript closely enough.

4. Common failure patterns:

   • No wake event: _match_wake_word didn't find the wake phrase in the scripted transcript. Check that config.wake_words matches exactly what's in the script (case-insensitive, but prefix-boundary aware).
   • No matched event after wake: The command transcript didn't score above matching_threshold (default 70). The scripted transcript must be close enough to a trigger's commands list.
   • Reply window: only 1 matched instead of 2: capture_transcript didn't receive the reply transcript. Check that ScriptedBackend.recreate() is a no-op and the FIFO queue still has the reply queued.
   • Test hangs past timeout: stop_event not set after last transcript. Check ScriptedBackend sets stop_event after the last popleft().