relive（relive 音频）

Re:Live - AI Clone Agent

1. Overview

Re:Live replicates a person as a digital clone: personality (chat logs → profile.md), voice (reference audio + CosyVoice3), and optionally appearance (video first frame). Output can be text / voice / video. Dialogue is persisted under the character directory and used in dual-track RAG. Execution: run python main.py <JSON_file_path> from this skill’s root directory. See README.md for details.

Environment (required): Always run python main.py ... inside a virtual environment created in this skill directory and install dependencies there, especially for voice / video synthesis. Typical setup (from workspace/skills/relive):

# Windows (PowerShell):
.\.venv\Scripts\Activate.ps1
# Linux/macOS:
source .venv/bin/activate

Quick start (when character already exists)

When the user says "talk to Martha" or uses /relive:Martha:

1. Read personality: Read .openclaw/workspace/skills/relive/storage/default_Martha/profile.md as the basis for reply style.
2. Single-turn dialogue (from skill root, e.g. .openclaw/workspace/skills/relive):

- When history is needed: write get_context.json (with user_id, target_id, content = user’s message) → python main.py get_context.json, use the returned context to help generate. - Main Agent generates reply text in character. - Write synthesize.json (content = reply text, user_message = user message, output_mode = text/voice/video) → python main.py synthesize.json to persist and optionally output voice/video.

New character: See "2. Creating a new character" below; order is init → upload → export_md → personality analysis → write profile.md → add to USER.md.

2. Creating a new character

2.1 When to create

When the user expresses intent like "clone/replicate someone", "create a digital twin of a deceased relative", or "create an AI persona from chat logs and voice", start the create new character flow.

2.2 Materials and directories

• Chat logs: User must upload (JSON supported, e.g. QQ/WhatsApp export). If only screenshots exist, parse them yourself; this skill does not. Confirm with the user which side is the character to clone.
• Reference audio: Put files in storage/{user_id}_{target_id}/voice_profile/ and must ask the user for the transcript of that audio; save it as corresponding.txt in the same directory.
• Reference image (optional): Create reference_image_url.txt under the character directory with one line, public URL, for output_mode: "video".

Character root: storage/{user_id}_{target_id}/ (e.g. storage/default_Martha/).

2.3 Steps (in order)

Run from the skill root directory. If the user does not provide chat logs, you can ask for character traits and go straight to step 4.

Step 1: Initialize directories

{ "type": "init", "user_id": "default", "target_id": "Martha" }

Run: python main.py init.json (or write to init.json then run; same below).

Step 2: Upload chat logs

{
  "type": "upload",
  "user_id": "default",
  "target_id": "Martha",
  "file_path": "/absolute/or/relative/path/filename.json",
  "file_type": "json",
  "self_name": "Jonas",
  "target_name": "Martha"
}

self_name / target_name must exactly match sender names in the chat log. Run: python main.py upload.json.

Step 3: Export to Markdown

{ "type": "export_md", "user_id": "default", "target_id": "Martha" }

Run: python main.py export_md.json to produce storage/default_Martha/chat.md.

Step 4: Personality analysis and profile.md

• Read storage/{user_id}_{target_id}/chat.md (split if too large).
• Use the LLM to analyze personality, catchphrases, how they address people, and style; write the result to profile.md in that directory (this step is done by the main Agent; there is no separate API).

2.4 After creation: update USER.md

Must add the character to the workspace root USER.md. Create (or extend) a section like "Re:Live characters", and:

• Add one bullet per character with a short English description, for example:

- **Martha**: pharmacy / biology student, gentle and friendly, loves mystery movies - **Dabao**: colloquial, warm and reliable friend who enjoys cooking and traditional activities

• Add a short instruction line that explains how to use this skill, e.g.:

Re:Live digital-clone skill. When the user types \ elive:<character_name>, read SKILL.md under .openclaw\workspace\skills\ elive\ for acting rules, and read profile.md under .openclaw\workspace\skills\ elive\storage\default_{target_name} as the persona definition for that character.

The main Agent reads this section from USER.md to know which Re:Live characters exist, how to describe them briefly, and how to route \ elive commands to this skill and the corresponding profile.md.

3. After character exists: entering the character quickly

3.1 Commands and state

• /relive:<target_id> (e.g. /relive:Martha): Enter relive mode; main Agent stores target_id in session state; subsequent messages to this skill use that id; replies are generated via relive and persisted under storage/{user_id}_{target_id}/.
• /relive:end: Exit relive and clear current character state.

As long as a current relive character exists, the flow is: read profile → get_context when needed → LLM generate → synthesize → persist.

3.2 Always read profile.md when entering character

Before each conversation with that character, must read storage/{user_id}_{target_id}/profile.md and inject it into the main Agent’s system/context so reply style is consistent.

3.3 Single-turn flow (three steps)

For each user message, from skill root:

1. get_context when needed: If history is needed, write get_context.json (content = user’s message), run python main.py get_context.json, use returned context to help generate.
2. Generate reply: Main Agent generates reply text in character.
3. synthesize to persist: Write synthesize.json (content = reply text, user_message = user message, output_mode = text/voice/video), run python main.py synthesize.json. Even for text-only, run synthesize if you want the turn in runtime and RAG (output_mode can be text or omitted).

3.4 Output modes

• text: Text only.
• voice: Text + voice clone (requires voice_profile + corresponding.txt). Environment: install deps and models per README; recommended to use a virtual environment (see README § Installation and Voice Models).
• video: Video generation API (Seedance, etc.); same synthesize entry with output_mode: "video". After video_task_id is returned, run the auto-generated video_wait.json to poll and download.

4. API parameters summary

• upload: self_name / target_name must exactly match sender names in the chat log.
• synthesize: output_mode optional text (default), voice, video; optional reference_image_url (if not passed, read from reference_image_url.txt under character directory); optional video_wait: true to poll in-call until video is done. After video success, video_wait.json is auto-generated; run python main.py video_wait.json to poll and download.
• video_generation_wait: type in JSON must be video_generation_wait; conventional filename is video_wait.json. Optional poll_interval_seconds, poll_timeout_seconds.

5. Notes

• Privacy: User data is isolated per character and used only for the current clone task.
• Ethics: Do not use for deception, forgery, or other misuse.

6. More reference (see README.md)

• Data storage layout: source/, chat.md, profile.md, runtime/, vector_db/, voice_profile/, cache/ under character directory → README.md#file-structure.
• Chat log formats: QQ/WhatsApp JSON format and field requirements → README.md#preparing-chat-logs.
• Voice output: Dependencies and model setup (recommended: use a virtual environment), corresponding.txt requirements → README.md#installation-and-voice-models-required-before-voice-output, README.md#advanced-features.
• Video generation: Config, polling, and video_generation_wait example → README.md#advanced-features.
• RAG: Dual-track retrieval, BM25, thresholds and testing → README.md#where-rag-searches.
• Troubleshooting: Message attribution, format failure, voice/video not generating, 404, etc. → README.md#troubleshooting.
• Core modules (implementation reference): main.py, orchestrator, importer, memory, engines → README.md#core-modules-implementation-reference.

Note (OpenClaw exec timeout): When this skill is invoked via OpenClaw’s exec, long CosyVoice3 voice synthesis may be killed by the default execution timeout. If you observe the process exiting with code 1 shortly after logging synthesis text ... and no audio file is written, check npm/node_modules/openclaw/dist/auth-profiles-*.js and increase DEFAULT_EXEC_TIMEOUT_MS (for example, from 5e3 to 180e3) so that long-running voice synthesis can finish.