dsh-dafeiyu-emotion 0.3.2 → 0.3.3

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## 0.3.3 — 2026-11-07
+
+- Added `petReactionEnabled` (default on): while a spoken reply plays, the local `dsh-dafeiyu` pet enters its thinking animation; when playback ends it plays the success celebration. The bridge uses a fake session id on the public session event bus, never mutates the real session, auto-cleans after 2.6 s and has a 60 s safety end.
+- Added `petSayEnabled` (default on): reply text is typed into the pet bubble through the local `dsh-dafeiyu` HelperProcess bridge (`state` THINKING, repeated `task` messages, final `pulse` SUCCESS). The emotion bundle mounts before the pet so a `HelperProcess.start` hook can capture the bridge; old installs need a reinstall/restart. This bypasses the reducer's session selection and multi-task card. It is bubble text, not a new speaking animation — true lip-sync still needs new pet art/animation frames.
+- Pet reaction lifecycle is now idempotent: duplicate `end` is ignored, `end` still works after the feature is disabled, the safety end is 180 s, unmount/session switch sends the end, and host teardown sends IDLE to any active bridge.
+- Fixed `/context?session=` returning another session's last analysis, the 3-sample trend comparing against zero, saving baseline-adjusted energy/pause as the next baseline, emotion-adaptive TTS falling back to neutral after the post-send reset, the global (not per-session) `ttsOnlyAfterVoice` window, bare-command `ttsReady` false positives, and the over-broad privacy cleanup prefix.
+- Added `POST /plugins/dsh-dafeiyu-emotion/pet/reaction` (`start` / `end`) for the client-side TTS lifecycle. `start` may carry the reply text for the local pet bubble; audio is never sent, and `end` sends only the phase/session id.
+- Fixed a stale TTS failure falling back to browser speech after a newer reply started, which could end the newer pet reaction early.
+- Added a per-result emotion evidence card: speech rate, pause ratio, pitch range, energy, text-vs-prosody source, confidence and the low-confidence ask-first note.
+- Added a one-click personal baseline from the evidence card, stored locally and sent with later analyses; the lightweight analyzer now normalizes pitch, energy, pause and pitch variation against that baseline.
+- Added a per-session emotion trend (`更紧绷` / `更放松` / `更积极` / `更低落`) after three analyses, exposed in the analyze response, the evidence card and `/context`.
+- `/context` now accepts `?session=` and returns `history` plus `trend`; `GET /events` also emits `kind: "context"` for other plugins.
+- `buildPrompt` now adds an explicit “do not assert, confirm first” clause when confidence is below 0.4.
+- Added emotion-adaptive TTS (`ttsEmotionAdapt`, default on): the reply voice nudges rate and pitch from the last detected emotion — brighter for happy, slower and lower for sad.
+- Added a privacy-panel cleanup action and `POST /plugins/dsh-dafeiyu-emotion/privacy/cleanup` to remove orphaned plugin temp directories older than ten minutes.
+- Fixed partial configs disabling TTS even though the schema default is `true`; a missing `ttsEnabled` now keeps the documented local-Piper default.
+- Fixed a disabled plugin still broadcasting assistant text to the spoken-reply stream when TTS was left enabled.
+- Fixed assistant events without a session id leaking into every open DSH window; unscoped events are now ignored.
+- Fixed bare `edge-tts` / `piper` commands being reported unavailable by `/health` just because `existsSync()` cannot see PATH.
+- Fixed a host that returns `false` from `submit()` being treated as a successful delivery; the result now offers retry instead of hiding the send button.
+
 ## 0.3.2
 
 - Added `ttsOnlyAfterVoice` (default `true`): spoken replies are limited to assistant messages within two minutes of a voice input, so normal coding chats do not read every answer aloud.

package/README.md

@@ -73,10 +73,16 @@
 - 🎛️ DSH 设置页里的完整设置卡片：启用、语言、增强后端、超时和录音上限；
 - 🧭 安装 Better Sidebar 后，可把「情绪耳朵」固定到右侧边栏；
 - 🎭 共情策略四档：温柔、直接、幽默、简洁；
+- 🧭 情绪证据卡：展开查看语速、停顿、音高、能量和文本线索，置信度低时明确提示“先追问”；
+- 🎯 个人基线：把一次平静说话设为基线，轻量模式按你的个人音高/能量归一化；
 - 📈 多段语音的情绪时间线，可悬停查看每段文本、情绪和置信度；
+- 📈 会话情绪趋势：连续几次语音后给出“更紧绷 / 更放松 / 更积极”等走向；
 - 🔐 设置页内的隐私与运行状态：最近音频大小、模型加载、worker PID、临时文件清理；
-- 🔌 本机 `/context` 接口，其他插件可以读取最近一次情绪上下文；
+- 🔌 本机 `/context` 接口，其他插件可以读取最近一次情绪上下文、趋势和最近记录；
 - 🔊 大肥鱼回复语音：Piper 本地模型默认朗读，Edge / 浏览器语音可选；
+- 🎚️ 情绪自适应语音：按上一句情绪微调语速和音调，喜悦更亮、低落更慢；
+- 🐋 桌宠回复互动：朗读时让桌面大肥鱼进入思考动画，气泡逐字显示回复，读完后庆祝；需要安装 `dsh-dafeiyu` 0.1.23+；
+- 🧹 隐私面板一键清理：删除 10 分钟前遗留的临时音频目录；
 - 🐋 大肥鱼主题的“鱼耳”入口，和普通对话共用同一个 composer；
 - 🧠 情绪标签：喜悦 / 低落 / 生气 / 紧张 / 惊讶 / 厌烦 / 平静；
 - 📈 额外给出 confidence、valence、arousal 和语气摘要；
@@ -86,6 +92,8 @@
 - 🧩 大肥鱼桌面角色由 `dsh-dafeiyu` 提供，本插件负责“听觉与情绪”；
 - 🛠️ 如果情绪分析失败，会保留明确错误，不会假装成功。
 
+> 桌宠互动走 `dsh-dafeiyu` 的 HelperProcess 协议：`state` 思考、`task` 逐字气泡、`pulse` 庆祝。新版 bundle 把情感插件排在 `dsh-dafeiyu` 之前以安装桥接钩子；旧安装需要重装并重启 DSH。嘴型同步仍需要新的角色动画素材。
+
 ## 安装
 
 ### 通过 1024Store
@@ -213,6 +221,9 @@ python -m pip install piper-tts
 | `language` | `zh` | 识别语言：`zh` / `en` / `auto` |
 | `empathyStyle` | `gentle` | 共情策略：`gentle` / `direct` / `humorous` / `concise` |
 | `ttsEnabled` | `true` | 是否朗读大肥鱼回复 |
+| `ttsEmotionAdapt` | `true` | 按上一句情绪微调语速和音调 |
+| `petReactionEnabled` | `true` | 朗读时驱动本地大肥鱼思考/庆祝 |
+| `petSayEnabled` | `true` | 朗读时把回复文本逐字显示在大肥鱼气泡里 |
 | `ttsOnlyAfterVoice` | `true` | 只朗读录音后两分钟内的回复 |
 | `ttsEngine` | `piper` | `piper` 本地 / `edge` 在线 / `browser` 系统语音 |
 | `ttsVoice` | `zh-CN-XiaoyiNeural` | Edge TTS 音色 |

package/cordis.patch.yml

@@ -1,4 +1,29 @@
 - insert:
+    # Emotion mounts before the pet so its top-level HelperProcess.start hook
+    # is in place when dsh-dafeiyu creates the desktop bridge.
+    - id: dsh-dafeiyu-emotion
+      name: dsh-dafeiyu-emotion
+      config:
+        enabled: true
+        backendRoot: ""
+        pythonPath: ""
+        language: zh
+        empathyStyle: gentle
+        ttsEnabled: true
+        ttsEmotionAdapt: true
+        petReactionEnabled: true
+        petSayEnabled: true
+        ttsOnlyAfterVoice: true
+        ttsEngine: piper
+        ttsVoice: zh-CN-XiaoyiNeural
+        ttsRate: "+6%"
+        ttsPitch: "+12Hz"
+        ttsCommand: ""
+        ttsPiperModel: ""
+        ttsPiperCommand: ""
+        maxAudioBytes: 12582912
+        startupTimeoutMs: 180000
+        requestTimeoutMs: 180000
     - id: dsh-dafeiyu
       name: dsh-dafeiyu
       config:
@@ -21,23 +46,3 @@
         clientTarget: auto
         balanceEnabled: true
         balanceSource: auto
-    - id: dsh-dafeiyu-emotion
-      name: dsh-dafeiyu-emotion
-      config:
-        enabled: true
-        backendRoot: ""
-        pythonPath: ""
-        language: zh
-        empathyStyle: gentle
-        ttsEnabled: true
-        ttsOnlyAfterVoice: true
-        ttsEngine: piper
-        ttsVoice: zh-CN-XiaoyiNeural
-        ttsRate: "+6%"
-        ttsPitch: "+12Hz"
-        ttsCommand: ""
-        ttsPiperModel: ""
-        ttsPiperCommand: ""
-        maxAudioBytes: 12582912
-        startupTimeoutMs: 180000
-        requestTimeoutMs: 180000

package/docs/MARKETPLACE.md

@@ -41,9 +41,12 @@ https://github.com/LiDu-china/dsh-dafeiyu-emotion/releases/download/v0.2.3/dsh-d
 3. 优先复用 DSH 自带 Voice Input/SenseVoice 做本地转写；
 4. 浏览器提取语速、停顿、基频、音高变化、能量等声学特征；
 5. 可选中英文情绪词表与声学原型融合，输出情绪、置信度、愉悦度、唤醒度；
-6. 生成“先承接情绪、再回答问题”的受控 prompt；
+6. 生成“先承接情绪、再回答问题”的受控 prompt；可展开情绪证据卡，置信度低时先追问确认；
 7. 自动填入并发送到当前会话；
-8. 如果本机部署了 VoiceSense，可切换到 SenseVoice + emotion2vec 增强模式。
+8. 可用一次平静说话设置个人基线，轻量模式按个人音高/能量归一化；
+9. 连续几次语音后生成会话趋势，并通过 `/context` 与 `kind:"context"` SSE 提供给其他插件；
+10. 朗读回复时通过本地 HelperProcess 协议驱动 `dsh-dafeiyu`：思考 → 逐字气泡 → 庆祝；音频从不发送；
+11. 如果本机部署了 VoiceSense，可切换到 SenseVoice + emotion2vec 增强模式。
 
 插件不常驻监听、不读屏幕、不声称读心。音频只进入本机临时目录，分析结束后删除。
 
@@ -65,6 +68,7 @@ https://github.com/LiDu-china/dsh-dafeiyu-emotion/releases/download/v0.2.3/dsh-d
 - Microphone permission
 - Lightweight mode: enable DSH Voice Input and prepare its local model
 - Enhanced mode: a local VoiceSense checkout with `.venv-ser` and cached models
+- Optional pet reaction: `dsh-dafeiyu` 0.1.23+ and the emotion plugin mounted first (the bundled patch does this; reinstall/restart old installs)
 
 ## Install
 
@@ -78,6 +82,8 @@ Published as `dsh-dafeiyu-emotion@0.3.2` on npm. The 1024Store entry refreshes i
 
 No audio upload. No screen capture. No persistent microphone. Temporary WAV is deleted after analysis.
 
+The optional pet reaction sends `start` / `end`, a session id, and (on `start`) the reply text to the local host so the pet can type it into its bubble. Audio is never sent to the pet, and the text is never uploaded to a cloud service.
+
 ## Maintainer
 
 [@LiDu-china](https://github.com/LiDu-china)

package/docs/RELEASE-0.3.3.md

@@ -0,0 +1,43 @@
+## 0.3.3
+
+大肥鱼互动与情绪上下文修复版。
+
+### 大肥鱼气泡互动
+
+- 新增 `petSayEnabled`，默认开启：助手语音回复播放时，本地大肥鱼气泡会逐字显示 `正在说：<回复文本>`。
+- 通过 `dsh-dafeiyu` 的 HelperProcess bridge 直接发送 `state` / `task` / `pulse` 协议消息，不再伪造会话，也不会在真实会话里留下多任务卡。
+- 回复开始时大肥鱼进入思考动画，结束时播放成功庆祝和音效。
+- 新增 `petReactionEnabled`，默认开启；可在设置里单独关闭动画或气泡文本。
+- `/health` 新增 `petBridgeReady`，用于确认 HelperProcess bridge 是否已挂上。
+- 情绪插件现在排在 `dsh-dafeiyu` 之前挂载，确保能捕获桌宠启动时创建的 bridge。
+
+> 从旧版本升级后必须完全退出并重启 the agent harness，再确认 `/health` 里 `petBridgeReady: true`。当前是思考动画、逐字气泡和结束庆祝，不是嘴型同步；真正的口型动画仍需要新的桌宠帧或协议。
+
+### 情绪上下文与稳定性
+
+- 修复 `/context?session=` 在未知会话时错误返回其他会话 `lastAnalysis` 的问题。
+- 情绪趋势改为至少 4 个样本才计算，避免 3 个样本拿 0 当基线。
+- 个人基线保存原始 `calibration` 特征，避免下一轮重复归一化。
+- 情绪自适应 TTS 不再在发送后重置为 neutral，会保留最近 120 秒的情绪。
+- `ttsOnlyAfterVoice` 改为按会话独立计时。
+- 裸 `ttsReady` 命令改用 `where` / `which` 探测并缓存，避免 PATH 命令误判。
+- 隐私清理只匹配插件自己的 `dsh-dafeiyu-emotion-*` 和 `dsh-dafeiyu-tts-*` 临时目录。
+- 修复音频 blob URL 在卸载或失败时未释放的问题。
+- 证据卡不再把缺失值显示成 `0`。
+- 修复 TTS 失败回退可能提前结束新回复互动的问题。
+- 桌宠互动 `end` 幂等；关闭功能、切换会话、客户端卸载和宿主 teardown 都会正确收尾。
+
+### 验证
+
+- `npm test`：11 个测试文件全部通过。
+- `npm run check`：通过。
+- `python -m py_compile backend/voice_worker.py`：通过。
+- `npm pack --dry-run`：发布包不包含测试与演示素材。
+
+### 安装 / 升级
+
+```bash
+dsh plugin --profile desktop add dsh-dafeiyu-emotion
+```
+
+升级后请完全退出并重新启动 the agent harness。