more fixes for testing

This commit is contained in:
Jacob Dubin
2026-04-16 07:18:33 -05:00
parent b1069805bf
commit 500b54a6b6
8 changed files with 292 additions and 19 deletions

View File

@@ -74,6 +74,7 @@ The current .NET pass covers only a narrow, explicitly synthetic subset of obser
- `EOS` emission after completed turns
- delayed `SKILL_ACTION` emission after `EOS` on completed turn flows to better match the Node oracle timing
- first richer vertical slice for joke/chat `SKILL_ACTION` playback
- fixture-backed joke-turn payload fidelity for `CLIENT_ASR -> LISTEN -> EOS -> delayed SKILL_ACTION`, including Node-like `EOS` envelope fields and the currently observed joke `SKILL_ACTION` metadata shape
This does not yet mean parity for:
@@ -81,8 +82,32 @@ This does not yet mean parity for:
- real STT provider integration and external ASR lifecycle timing
- early-EOS behavior
- multi-step skill lifecycles beyond the current synthetic playback response
- broad `SKILL_ACTION` payload coverage outside the currently observed joke/chat playback slice
- broader interaction, animation, or ESML command families
### Successful Joke Turn: What Is Grounded Now
The highest-confidence websocket vertical slice after the starter parity pass is now:
- inbound `CLIENT_ASR` carrying `"tell me a joke"`
- outbound synthetic `LISTEN` result with joke intent and remembered rules
- outbound `EOS` carrying `ts`, `msgID`, `transID`, and an empty `data` object
- outbound `SKILL_ACTION` about 75 ms later
- joke `SKILL_ACTION` payload shape aligned with the Node oracle for:
- `data.skill.id = "@be/joke"`
- `data.action.config.jcp.type = "SLIM"`
- `data.action.config.jcp.config.play.meta.prompt_id = "RUNTIME_PROMPT"`
- `data.action.config.jcp.config.play.meta.prompt_sub_category = "AN"`
- `data.action.config.jcp.config.play.meta.mim_id = "runtime-joke"`
- `data.action.config.jcp.config.play.meta.mim_type = "announcement"`
What remains intentionally unclaimed for that slice:
- whether the joke payload is complete beyond those fields
- whether other successful skills use the same payload shape
- whether additional websocket messages appear in other successful skill paths
- whether any timing gaps besides the observed 75 ms `EOS -> SKILL_ACTION` delay matter
Current raw-audio fallback behavior remains explicitly synthetic:
- when a buffered-audio turn can be resolved through the synthetic transcript-hint seam, `.NET` now auto-finalizes and emits `LISTEN` + `EOS` + `SKILL_ACTION`