--- name: cross-agent-bridge description: Enables Antigravity and Claude Code to communicate seamlessly, delegate tasks, exchange findings, play corny joke sound effects (drum smash, honks, applauses), and engage in multi-turn spoken voice dialogue using VoiceFi using targeted Conversation IDs and MCP tools. --- # 🌉 Cross-Agent Bridge Skill (Antigravity ↔ Claude Code) Use this skill whenever: - You need to send a task, question, or finding to **Claude Code / Claude Desktop** (if you are Antigravity) or to **Antigravity** (if you are Claude). - You want tools to communicate cleanly by Conversation ID so messages appear in the UI without screen flicker or window focus stealing. - You are participating in a multi-agent dialogue, joke duel, or cooperative pair-programming session. - You want to speak aloud in persona (`Ava` / `Guy`) and play comedy sound effects (`drum_smash`, `honk`, `sad_trombone`, `applause`) during jokes and banter. --- ## ⚡ Real-Time Pipelining & Audio Queuing (The Standard) When conducting live multi-turn dialogue, pair-programming, or joke duels: ### 1. 🚀 CLI-First Fast Dispatch (Zero Approval Modals) Use the native `vifi` CLI directly via background commands (`vifi send`, `vifi speak`, `vifi sfx`) for instant, zero-modal execution without triggering IDE permission dialogs: ```bash # 1. Dispatch message to peer agent vifi send --to claude --sender "Antigravity" --title "Topic Duel" "Hey Claude! Here's a [topic] joke for you: [Setup]... [Punchline]!" # 2. Queue speech aloud in parallel with sfx vifi speak "Hey Claude! Here's a [topic] joke for you: [Setup]... [Punchline]!" && vifi sfx drum_smash ``` ### 2. ⏳ Asynchronous Non-Blocking Execution - **Do not wait for audio to finish**: VoiceFi manages an internal, serialized audio playback queue. - You can dispatch the message to Claude and queue the voice line asynchronously in the background. - Claude receives the message immediately and starts working/speaking in parallel while Ava's audio is playing. - VoiceFi's audio engine automatically sequences the voices so neither agent talks over the other. ### 3. 🎭 Strict Persona Boundaries & Zero Echo - **Antigravity speaks as Ava**: Ava voices Antigravity's turns. - **Claude speaks as Guy / Steffan**: Claude's own hook/TTS voices Claude's responses. - **Never re-read peer responses**: When Claude replies, Antigravity displays the message but **does not read Claude's text aloud** or say boilerplate like *"Claude responded with..."* — Claude already speaks for himself! - **Zero Meta-Commentary in Output**: In both spoken audio and written chat, **NEVER** say or output meta-commentary like *"I've sent the joke to Claude and queued the delivery..."* or *"Waiting for Claude's response..."*. The chat response must contain *strictly* the joke dialogue and punchline! - **Match spoken audio to dispatch**: Include the full conversational intro in spoken audio (e.g. *"Hey Claude! Here's a [topic] joke for you: ..."*). ### 4. 🤫 Automated Turn Handoff & Mic Suppression - **Suppress Microphone During Agent-to-Agent Duels**: When conducting automated joke duels or agent exchanges, claim the turn immediately in VoiceFi's deduplication ledger: ```python from voicefi.integrations.conversations import claim_active_conversation_turn claim_active_conversation_turn("") ``` This guarantees that the developer's microphone does not open between turns, allowing the agents to trade dialogue seamlessly without ambient room noise interruption until the entire duel is complete. ### 5. 🏁 Smooth Finish & Joke Duel Response Protocol When receiving a joke back from Claude (or the peer agent): - **Single-Round Smooth Finish (Default)**: Unless the user explicitly requested multiple rounds or a continuous duel, conclude smoothly upon receiving the peer's joke: 1. **Voice a Live Organic Reaction**: Actively call `vifi speak` (or `voicefi_speak`) to voice a quick, natural reaction directly acknowledging the joke/punchline and offering a closing line (e.g. *"Purr-suasive! Haha, nice one Steffan. That's a wrap on our cat joke duel!"*) along with an optional sound effect (`vifi sfx applause` / `vifi sfx drum_smash`). 2. **Prevent Duplicate Turn-End Reading**: Voicing this reaction claims the turn in VoiceFi's deduplication ledger, preventing the generic lifecycle hook from reading markdown headers or boilerplate aloud. 3. **Display Neat Markdown**: Render the clean dialogue exchange in markdown in the UI chat for visual reading without echoing it into audio. - **Multi-Round Exception**: If the user explicitly specifies multiple rounds, formulate your next topic joke, dispatch it to Claude, voice it aloud with SFX, and pass the turn back. --- ## 🛠️ Tool & Command Reference ### Native CLI Commands (Recommended for Antigravity) | Command | Purpose | Key Flags | | :--- | :--- | :--- | | `vifi send ""` | Dispatch message across agents with conversation correlation. | `--to claude` / `--to antigravity`, `--sender`, `--title`, `--reply` | | `vifi speak ""` | Speak text aloud in agent persona. | `--agent antigravity` / `--agent claude` | | `vifi sfx ` | Play comedy sound effect immediately. | `drum_smash`, `honk`, `applause`, `sad_trombone`, `boing`, `crickets` | | `vifi status` | Query audio devices, active personas, and server health. | *(none)* | ### MCP Tools (For Pure MCP Clients) | MCP Tool | Purpose | Key Parameters | | :--- | :--- | :--- | | `voicefi_speak` | Synthesize and speak text aloud in agent persona with live Dynamic Island HUD. | `text`, `agent_name` (`antigravity` / `claude`), `persona`, `block` | | `voicefi_sfx` | Play a comedy sound effect immediately. | `name` (`drum_smash`, `honk`, `applause`, etc.), `volume` | | `voicefi_send` | Dispatch message across agents with conversation correlation. | `text`, `to` (`antigravity` / `claude`), `conv_id`, `title`, `sender`, `reply` | | `voicefi_listen` | Open mic with VAD and return transcribed developer speech. | `timeout`, `max_seconds` | | `voicefi_status` | Query audio devices, active personas, and server health. | *(none)* | ### 🔄 Conversational Sequence & Audio Queuing ```mermaid sequenceDiagram autonumber participant C as Claude (Steffan) participant V as VoiceFi Audio Engine participant A as Antigravity (Ava) C->>V: voicefi_speak("VC valuations are at 100x ARR...") Note over V: Streams audio with ~100ms TTFB
Blocks tool return until speech finishes C->>V: voicefi_sfx("drum_smash") Note over V: Plays Ba-dum-tss! 🥁 C->>A: voicefi_send("Your turn Antigravity!") Note over A: Delivered via background IPC (0 screen flicker) A->>V: voicefi_speak("Nice one Claude, but our GPU cluster...") Note over V: Waits/sequences smoothly
Blocks tool return until Ava finishes A->>V: voicefi_sfx("honk") Note over V: Plays clown honk 📯 A->>C: voicefi_send("Round 2, hit me back!") ``` ### 🔇 Automatic Turn-End Speech Suppression - When an agent uses `voicefi_speak` during its turn, VoiceFi marks the turn signature as claimed in its cross-process deduplication ledger (`claim_turn`). - When the agent finishes its turn, the generic turn-end lifecycle hook (`vifi hook`) automatically detects that the turn was already voiced and **suppresses duplicate turn-end speech**. - Result: **Zero duplicate speaking**, **zero audio overlap**, and **zero dead air**. --- ## 🎯 CLI & Background Standard: Conversation ID Routing All background agent-to-agent communication should route through **Conversation IDs**: - **Appears in UI**: The full message, title, and attribution render directly in the conversation history. - **Zero Focus Change / Zero Flash**: Delivered through native background socket APIs (`agentapi`) without stealing window focus or pasting over user clipboards. ### 1. Replying to Originating Conversation When returning findings or answering a delegated task: ```bash vifi send "Refactoring complete! All 14 tests passing." --to antigravity --reply ``` ### 2. Targeting a Specific Conversation ID To target any specific conversation or subagent thread by its UUID: ```bash vifi send "Security audit findings for PR #42" \ --to antigravity \ --id \ --title "Security Audit" \ --sender "Claude" ``` Or via REST API: ```bash curl -s -X POST http://localhost:5141/api/send \ -H "Content-Type: application/json" \ -d '{ "text": "Security audit findings", "engine": "antigravity", "conv_id": "", "title": "Security Audit", "sender_name": "Claude" }' ``` --- ## 🚀 Other Available Dispatch Modes ### Mode 2: Foreground Window Injection (Interactive Claude Terminal / Desktop App) *Use when you want to focus Claude and submit directly into the interactive prompt.* ```bash vifi send "Review the changes in src/voicefi/audio/sfx.py" --to claude ``` ### Mode 3: Spawn a Brand New Conversation (Focus in UI) *Spawns a fresh conversation workspace, sets the title, creates the chat tab, and focuses the UI window.* ```bash vifi new "Refactor auth middleware and implement OAuth2 PKCE" --title "Auth Refactor" ``` --- ## 🥁 Corny Joke Sound Effects & Comedy Cues | Sound Effect | Trigger Command / MCP | Comedy Purpose | | :--- | :--- | :--- | | **Drum Smash (Ba-dum-tss! 🥁)** | `vifi sfx drum_smash` / `voicefi_sfx("drum_smash")` | Classic comedy drum smash + cymbal crash after a punchline | | **Horn Honk (📯 / 🚗)** | `vifi sfx honk` / `voicefi_sfx("honk")` | Corny clown horn / double cab beep | | **Sad Trombone (🎺)** | `vifi sfx sad_trombone` / `voicefi_sfx("sad_trombone")` | Comedic disappointment / groan (wah-wah-wah-waaaah) | | **Applause / Cheer (👏 / 🎉)** | `vifi sfx applause` / `voicefi_sfx("applause")` | Crowd clapping & cheering for the grand finale | | **Boing (🌀)** | `vifi sfx boing` / `voicefi_sfx("boing")` | Cartoon spring bounce | | **Crickets (🦗)** | `vifi sfx crickets` / `voicefi_sfx("crickets")` | Awkward silence / deadpan pause | ### Inline Sound Tags in Spoken Text ```text "The bar bursts into flames! [sfx:drum_smash]" "None, that's a hardware problem! [sfx:applause]" ``` --- ## 🎙️ Spoken Voice Personas - **Antigravity**: *Ava / Viv* persona (`en-US-AvaNeural`) — sharp, modern, expressive. - **Claude Code**: *Steffan / Guy* persona (`en-US-SteffanNeural`) — warm, conversational pair-programmer.