JiboExperiments/OpenJibo/docs/regression-test-plan.md

# Regression Test Plan

## Purpose

This plan is the repeatable live regression checklist for OpenJibo Cloud releases.

Use [live-jibo-test-runbook.md](live-jibo-test-runbook.md) for the environment setup and capture mechanics. Use this file for what to test once the robot is connected and the hosted `.NET` cloud is running.

The goal is to reduce trial-and-error cycles: every live pass should prove the release theme, keep prior working paths warm, and produce enough evidence to separate payload bugs, local robot behavior, and STT quality issues.

## When To Run

Run this plan:

- after the last code change before calling a release complete
- after any fix that touches websocket turn finalization, local skill redirects, constrained yes/no, or STT
- before moving from `1.0.18` bug-fix closeout into `1.0.19` feature work
- after the Test 26 and Test 27 fixes, run at least the focused cloud-version, alarm/timer, photo/gallery, stop, volume, and blue-ring cleanup sections before deciding whether `1.0.18` is ready to freeze

For small feature slices, run the automated `.NET` tests plus the smoke checks and only the live sections that share the same machinery. Before release closeout, run the full current-release suite.

## Required Evidence

For each live pass, keep these artifacts together under a named test folder such as `artifact-output/jibo-test-N`:

- `.NET` console logs
- websocket captures and fixture exports
- HTTP captures when startup, update, backup, media, or upload paths are involved
- robot runtime logs pulled after the session
- operator notes with exact phrases attempted and visible robot/menu state

Record failures with the observed transcript, active listen rules, emitted websocket response shape, and whether the robot menu state agreed with the cloud response.

## Release Gates

A release is not ready until these are true or explicitly deferred in [development-plan.md](development-plan.md):

- focused `.NET` cloud tests pass
- running robot reports the expected cloud version by voice and `/health`
- `cloud version` settles without a self-listened `Cloudford` / generic chat tail
- no current-release path emits obsolete OpenJibo-only websocket events such as synthetic pending/context/ack packets
- known working live paths still work: startup, simple chat, radio, basic news, constrained yes/no, alarm, and gallery/create
- any remaining failure is classified as cloud payload, local robot state, STT/audio quality, environment/routing, or deferred feature gap

## Automated Baseline

Run before the live session:

```powershell
dotnet test tests\Jibo.Cloud.Tests\Jibo.Cloud.Tests.csproj --no-restore --nologo -v minimal
```

Expected result for the current baseline: all tests pass.

## Live Smoke Checks

Run these first so obvious environment problems do not pollute feature results:

1. Start the `.NET` cloud using the live runbook.
2. Confirm `/health` reports the expected version.
3. Confirm the robot is not in a local connection-lost state; if logs show `Q4-Server_connection_lost` or a fresh `jibo-server-service` reconnect, wait for it to clear before scoring voice behavior.
4. Ask `cloud version`; confirm Jibo speaks the same version and does not follow with `Cloudford`, `I heard...`, or another generic tail reply.
5. Run one simple chat turn.
6. Run one joke turn.
7. Confirm websocket capture is being written before continuing.

Stop and fix environment issues if startup, websocket connection, or capture output is not clean.

## Current `1.0.18` Regression Suite

### Radio

Goal: keep the local radio redirect path proven.

- Say `open the radio`.
- Say `play country music`.
- Expected: Jibo opens or resumes the radio locally, and the country phrase carries a `Country` station entity.
- Capture check: websocket output should be local `SKILL_REDIRECT` plus silent completion, not generic chat speech.

### News

Goal: keep the Nimbus-shaped cloud skill path proven.

- Say `tell me the news`.
- Expected: Jibo plays the current synthetic quick brief.
- Capture check: `LISTEN` match includes `cloudSkill = news`, followed by a `news` `SKILL_ACTION`.
- Current limitation: provider-backed and category-expanded headlines are deferred unless selected as the optional feature slice.

### Backup, OTA, And Share Yes/No

Goal: prove constrained yes/no prompts stay local and do not leak global launch rules.

- Trigger the update menu path when available and answer one short `yes` or `no` prompt.
- Exercise any available share/date/offer yes-no prompt and answer both `yes` and `no` across runs when practical.
- Observe backup-in-progress behavior separately from explicit voice commands.
- Do not treat a spoken `take a backup` failure as proof of the backup scheduler path; that command is not currently wired as a hosted-cloud voice feature.
- If the update menu reports backup-in-progress, record whether HTTP captures include any `Backup_*` targets; current evidence points to robot-local scheduler/status or log/upload load unless those calls appear.
- If Jibo announces backup-in-progress without update-menu interaction, note the local skill in robot logs; Tests 26 and 27 showed `@be/surprises-ota`.
- If the warning appears soon after startup or update, check for local `jibo-server-service` restart, notification reconnect, or `Q4-Server_connection_lost` before scoring it as a hosted backup defect.
- Expected: short `yes`/`no` replies map locally, empty replies no-input locally, and backup/download notifications are not repeatedly re-announced once acknowledged.
- Capture check: active rule remains the constrained rule such as `surprises-ota/want_to_download_now`, `settings/download_now_later`, `shared/yes_no`, or another stock prompt rule.

### Alarm

Goal: prove the clock skill behaves locally and menu state agrees after the `jibo test 24` fixes.

Start from a known state. If an alarm already exists, record it and clear it through the menu or a controlled voice delete before beginning.

Test these paths:

- explicit set: `set an alarm for 7:43 AM`, adjusted to a near-future time during the actual run
- compact set: `set alarm for 743`, adjusted to a near-future time during the actual run
- clarification: `set an alarm`, then answer the value prompt with a short time such as `7 44` or `7, 44`
- replacement: with an alarm already set, set a different alarm and answer the replacement prompt; verify whether the answer kept or replaced the old alarm
- value-prompt cancel: `set an alarm`, then say `cancel`
- voice delete: `delete my alarm` or `cancel alarm`
- voice delete variants from Test 26: `delete the alarm`, `delete alarm`, and, if ASR mishears it, record whether `delete along` maps to local clock delete
- no-input cleanup: allow one value prompt to miss or time out when practical
- timer sanity: `set a timer for 10 seconds`, let it fire or record the exact remaining state, then verify a second timer request does not report a stale already-running timer

Expected:

- successful set paths appear in the robot alarm menu and fire at the expected time
- replacement prompt answer changes or preserves the alarm consistently with the robot's question
- `cancel` inside the value prompt closes without scheduling
- voice delete clears the robot menu state
- local clock delete/cancel settles without generic chat speech or an open follow-up blue ring
- timer state agrees with what just happened on the robot; a reset gesture should not leave a phantom active timer in the next prompt
- empty value prompt turns complete locally instead of generic `I heard you` speech

Capture check:

- clock payloads use local `@be/clock` handoff with alarm entities when a value exists
- missing values stay in local clock clarification
- `CLIENT_NLU cancel` under `clock/alarm_set_value` or `clock/timer_set_value` maps to local clock `cancel`
- no-input under `clock/alarm_set_value` or `clock/timer_set_value` returns local `LISTEN`/`EOS` only

### Photo Gallery And Create

Goal: prove gallery/create no longer leaves stale listening state after yes/no or preview prompts.

Test these paths:

- `open photo gallery`
- if gallery is empty, answer `yes` to the offer to take a picture
- if the robot hears `open photogal` or another close gallery alias, verify it still launches gallery
- take one photo and answer the keeper prompt with `yes`
- repeat a gallery empty prompt or create keeper prompt with a missed/empty answer when practical
- if using disposable test photos, test delete confirmation once with `no` and once with `yes`

Expected:

- empty gallery `yes` redirects to `@be/create`
- empty gallery `no` exits cleanly when tested
- keeper `yes` completes and Jibo settles without a stale blue ring
- after gallery settles, context-only tails do not produce delayed generic replies such as `that's` or `I didn't hear you`
- transcript-bearing `yes` under gallery `shared/yes_no` is consumed even when the robot reports `@be/gallery` context
- empty `shared/yes_no`, `create/is_it_a_keeper`, and `gallery/gallery_preview` turns no-input locally instead of generic `I heard you`
- delete confirmation only deletes on a positive `yes`

Capture check:

- gallery launch redirects to `@be/gallery`
- create photo redirects to `@be/create/createOnePhoto`
- local no-input replies keep the active constrained rule and strip unrelated global launch rules
- active `shared/yes_no` is not suppressed merely because the current context is `@be/gallery`
- post-gallery binary audio does not continue buffering unless a fresh `LISTEN` appears

### STT And Audio Quality

Goal: avoid misclassifying transcript failures as payload regressions.

For every failed voice turn, record:

- phrase attempted
- transcript observed in websocket capture
- active listen rule
- whether the transcript was empty, collapsed, or semantically wrong
- whether local `ffmpeg` or `whisper.cpp` logged an error

Expected:

- no `ffmpeg` failure should become the dominant failure mode for non-Opus buffered audio
- short replies such as `yes`, `no`, `cancel`, and short alarm times should either map correctly or be classified as STT misses with evidence

### Stop And Volume

Goal: prove the added lightweight device-control slice before closing `1.0.18`.

Test these phrases:

- `stop`
- `stop that`
- `never mind`
- `never mind.` or any punctuated transcript form observed in the capture
- `turn it up`
- `turn it down`
- `set volume to six`
- `set volume to 6`
- `show volume controls`

Expected:

- stop commands settle the robot locally without generic chat speech
- `turn it up` and `turn it down` adjust volume or at least produce the stock local volume event/log
- `set volume to six` sets or attempts to set the local volume level to `6`
- `show volume controls` opens the settings volume panel
- after `show volume controls`, the robot settles without a trailing `I heard you`

Capture check:

- stop emits `nlu.intent = stop`, `nlu.domain = global_commands`, then redirects to `@be/idle`
- punctuated `Never mind.` still maps to global stop, not generic chat
- relative volume emits `nlu.intent = volumeUp` or `volumeDown`, `nlu.domain = global_commands`, and `entities.volumeLevel = null`, with no `SKILL_ACTION` cloud speech
- absolute volume emits `nlu.intent = volumeToValue` and `entities.volumeLevel` matching the requested value, including the observed `Set Volume 2-6.` homophone shape, with no `SKILL_ACTION` cloud speech
- volume controls redirects to `@be/settings` with `nlu.intent = volumeQuery`
- passive `@be/settings` / `settings/volume_control` audio tails complete locally and do not reopen Nimbus fallback speech

### Blue-Ring Cleanup

Goal: catch the Test 26 no-`LISTEN` buffering regression and the Test 27 diagnostic speech-tail regression quickly.

- After any local skill redirect or generic chat reply, wait five to ten seconds before issuing the next phrase.
- If the blue ring remains open, record the active transID and whether the websocket capture shows a new `LISTEN`.
- After `cloud version`, wait five to ten seconds and confirm there is no fresh no-transcript hotphrase launch `LISTEN` that turns speech tail into generic chat.
- Expected: binary audio for an existing transID is ignored until a fresh valid `LISTEN` appears; blank hotphrase turns clear instead of buffering indefinitely; diagnostic speech tails do not reopen launch listens.
- Capture check: long-running context-only transactions should not accumulate buffered audio chunks or stay `AwaitingTurnCompletion = true`; a late ignored diagnostic `LISTEN` may appear as cleanup telemetry but should not set `SawListen` or buffer audio.

## Optional Feature Slice Checks

When a new feature is added before a release closes:

- add two or three exact phrases to this section before live testing
- capture one successful path and one near-miss phrase if the feature is voice-routed
- keep the test narrow enough that a failure can be fixed or deferred without reopening the whole release

For the current candidate list, add cases here when implemented:

- robot age/persona: `how old are you`

## After The Run

After each session:

1. Summarize pass/fail by section.
2. Mark each failure as cloud payload, local robot state, STT/audio, environment, or deferred gap.
3. Import any high-value websocket fixture.
4. Update [development-plan.md](development-plan.md) with latest live evidence.
5. Update [feature-backlog.md](feature-backlog.md) with what remains in the current release versus what moves to the next release.