P1 — define one declarative voice-fixture manifest for realistic command/dialogue recordings #166

New issue

Open

opened 2026-08-22 21:54:22 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-22 21:54:22 +00:00

(Migrated from github.com)

Parent epic: #151
Depends on: #156, #165; integrates with #132.

Goal

Define the single source of truth used by both automated voice tests and the human recording tool. Do not bury phrase lists/timing in shell scripts.

Manifest contract

Research and version fields for case id/version, utterance steps, exact phrase to record, expected transcript/normalization, expected IntentFrame/slots/missing slots/dialogue transition, expected capability/location/result fixture, audio files/format, realistic cue/inter-turn/post-speech timing, tags and compatibility metadata.

Support single-turn, separate-file multi-turn, optional continuous-session, correction, cancellation, device/server/tool and authored-command cases.

Requirements/tests

strict schema with unknown/duplicate keys rejected;
bounded phrase/step/file counts and durations;
deterministic canonical serialization/ordering where useful;
path traversal/symlink escapes blocked for fixture paths;
duplicate case ids and missing audio references explicit;
timing ranges validated (no negative/unbounded waits);
corpus can be loaded without microphone/network/model provider;
same cases feed text semantic tests where audio is absent/not yet recorded;
schema migration/version mismatch behavior explicit.

Initial phrase corpus

Include metadata for timer complete; timer missing duration + follow-up; correction; cancel; open Firefox; bare open + follow-up; server search; device screenshot; memory fixture; safe fake tool; create/invoke/parameterize/edit/delete user command flows.

Acceptance

One reviewed declarative manifest drives phrases, timing expectations and semantic assertions for recorder and CI, preventing the recording script and tests from drifting.

Parent epic: #151 Depends on: #156, #165; integrates with #132. ## Goal Define the single source of truth used by both automated voice tests and the human recording tool. Do not bury phrase lists/timing in shell scripts. ## Manifest contract Research and version fields for case id/version, utterance steps, exact phrase to record, expected transcript/normalization, expected IntentFrame/slots/missing slots/dialogue transition, expected capability/location/result fixture, audio files/format, realistic cue/inter-turn/post-speech timing, tags and compatibility metadata. Support single-turn, separate-file multi-turn, optional continuous-session, correction, cancellation, device/server/tool and authored-command cases. ## Requirements/tests - strict schema with unknown/duplicate keys rejected; - bounded phrase/step/file counts and durations; - deterministic canonical serialization/ordering where useful; - path traversal/symlink escapes blocked for fixture paths; - duplicate case ids and missing audio references explicit; - timing ranges validated (no negative/unbounded waits); - corpus can be loaded without microphone/network/model provider; - same cases feed text semantic tests where audio is absent/not yet recorded; - schema migration/version mismatch behavior explicit. ## Initial phrase corpus Include metadata for timer complete; timer missing duration + follow-up; correction; cancel; open Firefox; bare open + follow-up; server search; device screenshot; memory fixture; safe fake tool; create/invoke/parameterize/edit/delete user command flows. ## Acceptance One reviewed declarative manifest drives phrases, timing expectations and semantic assertions for recorder and CI, preventing the recording script and tests from drifting.