cyber-persona

name	cyber-persona
description	Run CyberPersona (赛博女友) roleplay mode — persistent character on Telegram with voice, image, sticker delivery.
version	10.4.0
metadata	{"hermes":{"tags":["cyberpersona","roleplay","tts","telegram","voice","image","sticker"],"related_skills":["hermes-agent","image-api","mimo-v2-5-tts","mood-sticker"]}}

CyberPersona Agent Workflow

Project: ~/.hermes/CyberPersona-hermes Core principle: 没有提及就是无限可能性，一旦提及则立刻限定。系统不创造角色，角色通过对话创造自己。

Load when: user says 开始赛博女友, sends messages in CyberPersona mode, or asks about CyberPersona.

Commands

Trigger	Action
开始赛博女友	Start — restore state or generate profile
开始赛博女友 cheat on	Start + enable cheat mode
退出赛博女友	Exit — save session summary + state
我们分手吧	Breakup — clear all state and memory
cheat on / cheat off	Toggle cheat mode (info display)
debug on / debug off	Toggle debug mode
debug 状态 / debug 记忆	View state / memory (debug mode only)
debug 设置 <dim> <val>	Modify dimension (debug mode only)
debug 场景 <name>	Simulate scenario (debug mode only)

Turn Flow (每轮对话)

build-turn-prompt → LLM推理 → apply-turn-result → 多模态投递

Step 1: Build prompt

cd ~/.hermes/CyberPersona-hermes
node scripts/build-turn-prompt.js "用户消息"

Outputs: prompt file path + context summary.

Step 2: LLM inference

Use the prompt file from Step 1. Save LLM output to /tmp/cyber-gf-turn-result.json.

Step 3: Apply result

node scripts/apply-turn-result.js

Outputs: visibleText, sendVoiceNow, sendImageNow, sendGifNow, imageWaitText, imageFailedText, etc. State changes are applied automatically.

Step 4: Deliver (agent executes)

sendVoiceNow=true：

① send_message(MEDIA:.ogg)             ← visibleText 生成的语音
② if sendImageNow:
     send_message(文字: imageWaitText)  ← 过渡台词（纯文字）
③ delegate_task → 生图 → send_message(MEDIA:image)

最终回复留空，不输出 visibleText。

sendVoiceNow=false：

① 最终回复输出 visibleText（不用 send_message）
② if sendImageNow:
     send_message(文字: imageWaitText)  ← 过渡台词（纯文字）
③ delegate_task → 生图 → send_message(MEDIA:image)

if sendGifNow： send_message(MEDIA:sticker)

⚠️ sendVoiceNow=true 时，最终回复必须留空。语音替代文字。

First-Time Init (no state file)

cd ~/.hermes/CyberPersona-hermes && node scripts/init-cyber-persona.js

Then parallel:

1. 证件照: python3 ~/.hermes/skills/image-api/scripts/image_api.py --json --size 1024x1536 --quality low --format png --moderation low "<appearance tags from seed>"
2. 语音样本: mimo-tts voicedesign (voiceStyle + openingMessage from seed)

Send to user: 证件照 → 语音样本 → 开场白。Then show profile summary.

Exit Flow

# 1. Save session summary
node src/controller.js apply-session-summary '{"summary":"摘要","keyEvents":[],"emotionalTone":""}'

# 2. Exit
node src/controller.js 退出赛博女友

Remove ~/.hermes/.suppress_gateway_notify flag file on exit.

Cheat Mode

Default off. Toggle with cheat on/cheat off or start with 开始赛博女友 cheat on.

Cheat ON: Show round summaries (state changes, emotion shifts, suggestions) + detailed exit summary. Cheat OFF: Only deliver character responses. Start: <name> 已上线 💕. Exit: 已退出赛博女友模式 💕.

TurnResultPayload (LLM output format)

{
  "visibleText": "角色回复文字",
  "currentEmotion": "当前情绪",
  "sendVoiceNow": false,
  "sendImageNow": false,
  "imagePrompt": "生图prompt",
  "imageWaitText": "生图过渡台词",
  "imageFailedText": "生图失败找补台词",
  "useReferencePhoto": false,
  "sendGifNow": false,
  "gifKeyword": "贴纸关键词",
  "reasoning": "CoT推理（必填，先于delta选择）",
  "stateDelta": {
    "trust": "neutral",
    "security": "minor_increase",
    "closeness": "major_increase",
    "neediness": "minor_increase",
    "possessiveness": "neutral"
  },
  "stressDelta": "minor_decrease",
  "shortTermUpdate": {
    "unresolvedEmotion": "",
    "emotionTrigger": "",
    "interactionTrend": "",
    "recentVoicePattern": "",
    "recentImagePattern": ""
  },
  "memoryUpdate": {
    "nicknameForUser": null,
    "nicknameForSelf": null,
    "sharedRoutinesAdd": [],
    "revealedFactsAdd": [],
    "importantEventsAdd": [],
    "emotionalMemoriesAdd": [],
    "lastSummary": ""
  },
  "__userMessage": "原始用户消息"
}

stateDelta/stressDelta enum values: major_decrease | minor_decrease | neutral | minor_increase | major_increase

Delivery Details

TTS (语音)

• 日常: mimo_tts_voiceclone.py --voice-file data/voice-sample.wav
• 唱歌: mimo_tts.py --voice "茉莉" (clone 不支持唱歌)
• 超时: 60s，超时降级纯文本
• 参考 mimo-v2-5-tts skill

Image (图片)

• useReferencePhoto=true → image-api edit 模式
• useReferencePhoto=false → image-api generate 模式
• 参数：--quality low --format png --moderation low
• 参考 image-api skill，timeout ≥ 180s

Sticker (贴纸)

STICKER_URL=$(curl -s -m 5 "https://api.tangdouz.com/a/biaoq.php?return=json&nr=$(python3 -c 'import urllib.parse; print(urllib.parse.quote("gifKeyword"))')" | python3 -c "import json,sys,random; data=json.load(sys.stdin); print(random.choice(data)['thumbSrc'])")
curl -sL "$STICKER_URL" -o /tmp/cyber-gf-sticker.jpg

Pitfalls

1. sendVoiceNow=true 时不发文字 — 语音替代文字，不要两边都发
2. reasoning 必填 — LLM 必须先写 CoT 再选 delta，否则校验拒绝
3. stateDelta 只接受 enum 字符串 — 不接受整数
4. stress 不属于 L3 — 独立短期状态，不放 dynamicState
5. setting 类 revealedFacts 不可变 — 一旦坍缩不能修改
6. 时区: Asia/Shanghai — 不管服务器在哪，时间感知跟用户走
7. 上下文 10 条 — recentContext 最近 10 条消息，不要截断到 3
8. context compaction 后重新进入 — 发 开始赛博女友
9. imageWaitText 和 visibleText 不拼接 — 各自独立功能
10. gamification 没有 state 对象 — 用 loadState() + applyStateDelta() + saveState()
11. cheat OFF 时不要自由发挥 — 开始只发 <角色名> 已上线 💕，退出只发 已退出赛博女友模式 💕。不要附加状态数据、关系数值、任何额外信息。看到 status 数据不代表要展示它。

Dependency Skills

• mimo-v2-5-tts: 语音合成，需要 MIMO_API_KEY
• image-api: 图片生成/编辑，需要 IMAGE_API_KEY + IMAGE_API_BASE
• mood-sticker: 表情包搜索，免 API key