sk-localprobe

Overview ¶

Command sk-localprobe verifies that the SpeechKit kernel libraries produce working Dictation, Assist, and Voice Agent results against LOCAL models only (Whisper.cpp + Gemma via llama-server). It is driven by install-e2e-windows.yml after a silent NSIS install to prove the post-install kernel works end-to-end with no cloud keys.

sk-localprobe does NOT need a running SpeechKit.exe. It imports the same kernel packages (internal/stt, internal/localllm, internal/voiceagent/cascaded) the Wails app uses and exercises them directly. This decouples the local-only guarantee from any Wails initialization path while still proving the production library code works against the installed binaries + models.

Required flags:

--whisper-model    path to ggml whisper .bin file
--gemma-model      path to Gemma .gguf file (or any GGUF the local
                   llama-server supports)
--fixtures         testdata/e2e root with dictation/, assist/,
                   voiceagent/ subdirs

Optional:

--report           JSON report path. Default sk-localprobe-report.json
--timeout          per-mode timeout. Default 5m (cold first-run can
                   take a few minutes when llama warms up).
--skip-voiceagent  skip the two-turn voice-agent scenario (debug)

Exit codes:

0 — all three modes passed
1 — at least one mode failed
2 — setup error (binary not found, model missing, etc.)