- keep websocket parity fixture-driven, starting with exact sequencing and payload-shape fidelity for the successful joke vertical slice before claiming broader skill coverage
The latest live test round tightened up three priorities:
- yes/no turns need explicit constrained follow-up handling instead of generic chat routing
- skill invocation still depends too much on narrow phrase matching and is vulnerable to STT drift
- local buffered-audio STT in `.NET` is useful for discovery, but it is not yet stable enough to be the default live-test assumption
Evidence from the latest `2026-04-18` captures:
- several buffered-audio turns never produced a usable transcript because the local `whisper.cpp` path was missing or the temporary normalized Ogg file was rejected by `ffmpeg`
- some recognized phrases fell into placeholder provider replies because the intent was recognized but the feature path behind it is still a stub
- short yes/no responses need the same session-aware treatment already prototyped in Node, especially for create-flow style follow-ups
Evidence from the latest word-of-the-day capture round:
- yes/no photo confirmation improved and now completes through the constrained follow-up path
-`CLIENT_NLU` menu navigation is surfacing richer `destination` entities such as `snapshot`, `fun`, and `word-of-the-day`
- word-of-the-day guesses can arrive as structured `CLIENT_NLU` turns with `intent=guess`, `rules=["word-of-the-day/puzzle"]`, and `entities.guess=<word>`
- those structured turns should be treated as first-class cloud inputs even when no free-form transcript is present
Evidence from the smaller `2026-04-18/19` hotphrase and word-of-the-day verification bundle:
- hotphrase silence can still auto-finalize into a generic `heyJibo` fallback, which sounds confused on-robot compared with a dedicated greeting path
- voice-triggered `loadMenu + destination=word-of-the-day` reaches Nimbus successfully, but Nimbus still expects a follow-up cloud skill response and times out if launch stops at `LISTEN` + `EOS`
- the newer `jibo test 2` bundle shows voice launch now reaches Nimbus and receives a cloud response, but a generic `SLIM/RUNTIME_PROMPT` just says "starting word of the day" instead of performing the menu-style redirect the on-screen path uses
- the `jibo test 3` bundle confirms Nimbus rejects `REDIRECT` in that cloud-skill slot, so the better next experiment is to hint the on-robot target skill directly on the synthetic `LISTEN` result and skip Nimbus `SKILL_ACTION` entirely for word-of-the-day launch
- the same bundle also shows `word-of-the-day/right_word` cleanup turns need a short ignore window for trailing audio or the robot can stay stuck in a blue-ring listening state
- the `jibo test 4` bundle exposed a broader websocket issue: inbound robot `LISTEN` setup packets were still being routed through turn finalization instead of just priming pending state, which can corrupt menu and word-of-the-day flows by treating setup turns like resolved intents
- the `jibo test 5` bundle suggests the remaining WOD launch and post-win cleanup bugs share the same root cause: we were leaving the robot-side `cloudSkillResponse` promise unresolved on `word_of_the_day`, `word_of_the_day_guess`, and `word-of-the-day/right_word`, so the latest .NET pass now emits a completion-only silent `SKILL_ACTION` for those paths instead of stopping at `LISTEN` + `EOS` or going fully silent
- the `jibo test 6` bundle plus the attached `@be` source snapshot refine that diagnosis: Nimbus does accept the silent completion response, but treats it as a normal `SLIM/RUNTIME_PROMPT` instead of a skill redirect, while the successful on-robot path is built around `menu + domain=word-of-the-day` skill switching through `SkillSwitchScheduler`
- the attached `be-framework.js` adds one more strong clue: the Be relaunch hook reads `skillData.nlu.skill`, so synthetic cloud launch turns for word-of-the-day should carry the explicit target skill name in the outbound NLU payload instead of expecting the robot to infer it from `intent/domain` alone
- the `JiboOs/V3.1` Nimbus source confirms the hotphrase/global launch path still routes through `@be/nimbus` and waits on `listenResult.cloudSkillResponse`, while Nimbus only supports a narrow set of cloud JCP behaviors and does not use cloud `REDIRECT` to jump into local skills; by contrast, the post-win `word-of-the-day/right_word` turn is a local `Optional-Response`, so the cleaner robot-side closeout is to synthesize an immediate empty `LISTEN + EOS` no-response result rather than replying with only `SKILL_ACTION`
- the same `jibo test 6` capture also shows the blue-ring cleanup loop was partly self-inflicted in `.NET`: after `word-of-the-day/right_word` we stopped the active turn, but later stray binary audio on the same transID could still re-arm buffering even without a fresh `LISTEN`, so the next pass now requires a real listen phase before post-turn audio can reopen buffered completion
- the local buffered-audio seam is still producing repeated `whisper.cpp returned no transcript` and `ffmpeg ... Codec not found` failures, so lightweight waveform or energy screening is worth considering once the core launch flow is stable
1. preserve and interpret yes/no turn constraints from observed listen rules
2. broaden phrase-to-intent matching for the small set of known working skills before moving to larger NLU ambitions
3. keep synthetic transcript hints as the most reliable parity path when captures already provide them
4. continue evaluating whether local preprocessing is worth further investment or whether managed STT should replace it for the next serious testing phase
5. start separating laptop-local capture storage from the eventual hosted retention/export path so group testing does not depend on repo-local zip handling
The current evidence in captures, fixtures, and Node behavior supports three main cloud interaction paths:
1. local Jibo behavior observed by the cloud
The robot or its local skill stack already interpreted the turn and the cloud mainly tracks, acknowledges, or lightly completes it.
2. local Jibo behavior overridden or redirected by the cloud
The robot reports the turn state, but the cloud chooses a different synthetic reply path.
3. raw audio interpreted by the cloud
The robot sends buffered audio and the cloud performs transcript resolution before sending back `LISTEN`, `EOS`, and ESML-driven playback.
Those are the right primary buckets for now. Additional side channels may still emerge later, especially around proactive traffic, direct skill/service sockets, or future on-device OS changes, but they should be treated as extensions to this model until captures prove otherwise.
-`word-of-the-day/right_word` closeout should not emit a synthetic `match`; otherwise Jetstream promotes it into `globalTurnResult` and Global Service relaunches Nimbus a few seconds later with a `Cloud Skill Response Timeout`.
- Voice `play word of the day` hotphrase launch still enters Global Service first, so a synthetic `LISTEN` result alone is not enough. The next-most-correct transport hint is a direct `SKILL_REDIRECT` event aimed at `@be/word-of-the-day`, alongside the menu-shaped `LISTEN` payload.
- Stock OS also keeps the original hotphrase/global launch cloud response promise alive even after the redirect succeeds, so voice WOD launch needs an explicit silent `SKILL_ACTION` completion on the same transID to avoid later cloud-response culling and an interrupted game state.
- Auto-dismissing `word-of-the-day/right_word` with a no-input `LISTEN`/`EOS` stops the listening ring, but it does not close the WOD UI by itself. Pairing that no-input closeout with an explicit redirect back to `@be/idle` is the current cleanest approximation.
- OTA/update yes-no prompts can advertise `$YESNO` only through ASR hints rather than `listenRules`, so short denials like `no` need to be recognized from `listenAsrHints` too.
- Spoken WOD guesses should preferentially snap to the closest offered hint when Whisper lands very close to one of the menu words, since near-misses like `haglet` for `aglet` are common in live testing.
- The stock robot still misroutes constrained local turns if the cloud echoes `globals/*` rules back on the reply. For spoken WOD guesses and settings/update `no`, we should only return the local rule (`word-of-the-day/puzzle`, `settings/download_now_later`, etc.) so Global Service does not relaunch Nimbus.
-`@be/radio` is a true local skill, not a cloud placeholder.
- Its `open(result, refresh, previousSkillName)` path treats `result.nlu.intent === "menu"` as a `play` launch.
-`result.nlu.entities.station` is the genre selector, and `Country` is a real supported station key from the robot's `genres.json`.
- The smallest stock-shaped cloud handoff for voice launch is therefore a local `SKILL_REDIRECT` to `@be/radio` with `nlu.intent = "menu"`, optional `entities.station`, and a silent completion to settle the hotphrase cloud response.
- Nimbus explicitly treats `match.cloudSkill === "news"` like the GQA path and waits on `cloudSkillResponse`.
- The first OpenJibo news pass should therefore use a real cloud-skill shape, not a generic placeholder chat reply.
- For now, the content can stay synthetic while the protocol is grounded: `match.cloudSkill = "news"` plus a supported `SLIM` announcement response is enough to validate the robot path before provider-backed headlines arrive later.
- The newest `.NET` pass now routes `open the clock` into the direct `askForTime` clock-view path, moves plain time/date/day questions onto stock-shaped local `@be/clock` handoffs, and keeps malformed timer/alarm requests on a clarification reply path instead of generic chat echo.
-`@be/gallery` is the local gallery browser and opens from `intent = "menu"`.
-`snapshot` and `photobooth` are not gallery submodes; stock main-menu logic remaps them into `@be/create` with `createOnePhoto` and `createSomePhotos`.
- The newest `.NET` pass keeps that routing, adds local-file persistence for media metadata, and serves stored media URLs back through `/media/{path}` as a first hosted-gallery slice.
- The remaining gap is binary fidelity: the current HTTP capture path stores request bodies as text, which is enough to preserve metadata and a placeholder payload, but may still be too lossy for perfect thumbnails/original fetches.
Latest update and state findings:
- unstaged update queries should not fabricate placeholder no-op manifests, because stock settings logic can treat any returned object like a pending update
- the hosted `.NET` cloud now persists update/media/backups state to a local state file by default, which is a better bridge toward Azure SQL / Blob storage than the old process-memory-only behavior
- full endpoint inventory beyond the current Node mapping
- OTA-driven recovery
- paid hosted plans or donation-supported hosting
- deeper on-device bridge and OS modernization
- more capable skill/runtime integration
- possible LLM or tool-use patterns inspired by workshop-era experimentation
## MCP-Like Ideas
Recent MIT workshop materials suggest experimentation around modern AI tooling for Jibo, including an MCP-oriented idea. We should treat that as inspiration for future OpenJibo directions, not as a present dependency or supported integration.
