476308d0df
Turn system redesign:
- Turn order auto-populated from config bindings (all bot accounts)
- No manual turnOrder config needed
- Humans (humanList) excluded from turn order automatically
- Dormant state: when all agents NO_REPLY in a cycle, currentSpeaker=null
- Reactivation: any new message wakes the system
- Human message → start from first in order
- Bot not in order → start from first
- Bot in order → next after sender
- Skip already-NO_REPLY'd agents when advancing
Identity injection:
- Group chat prompts now include agent identity
- Format: '你是 {name}（Discord 账号: {accountId}）'
Other:
- Remove turnOrder from ChannelPolicy (no longer configurable)
- Add TURN-WAKEUP-PROBLEM.md documenting the NO_REPLY wake-up challenge
- Update message_received to call onNewMessage with proper human detection
- Update message_sent to call onSpeakerDone with NO_REPLY tracking
109 lines
5.5 KiB
Markdown
109 lines
5.5 KiB
Markdown
# Turn-Based Speaking: Wakeup Problem
|
|
|
|
## Context
|
|
|
|
WhisperGate implements turn-based speaking for Discord group channels where multiple AI agents coexist. Only one agent (the "current speaker") is allowed to respond at a time. Others are silenced via a no-reply model override.
|
|
|
|
## The Problem
|
|
|
|
When the current speaker responds with **NO_REPLY** (decides the message is not relevant to them), the turn advances to the next agent. However, **the next agent has no trigger to start speaking**.
|
|
|
|
### Why This Happens
|
|
|
|
1. A message arrives in the Discord channel
|
|
2. OpenClaw routes it to **all** agent sessions in that channel simultaneously
|
|
3. The WhisperGate plugin intercepts at `before_model_resolve`:
|
|
- Current speaker → allowed to process
|
|
- Everyone else → forced to no-reply model (message is "consumed" silently)
|
|
4. Current speaker processes the message and returns NO_REPLY
|
|
5. `message_sent` hook detects NO_REPLY → turn advances to next agent
|
|
6. **But the next agent already "consumed" the message in step 3** — their session processed it (as no-reply) and moved on
|
|
7. No new message exists to trigger the next agent
|
|
|
|
### The Result
|
|
|
|
After a NO_REPLY, the next speaker sits idle until a **new** message arrives in the channel (from a human or another source). The original message that should have been passed to the next speaker is lost.
|
|
|
|
## When This Matters
|
|
|
|
- **Single-round conversation**: Human asks a question → Agent A says NO_REPLY → Agent B should answer but can't
|
|
- **Chain conversations**: Agent A defers → Agent B defers → Agent C should speak but never gets triggered
|
|
|
|
## When This Doesn't Matter
|
|
|
|
- **End-symbol responses**: When an agent actually speaks (ends with 🔚), the turn advances and the next agent will respond to the **next** message. This is fine.
|
|
- **Human-driven channels**: If humans keep sending messages, the dormant state resolves quickly.
|
|
|
|
## Possible Solutions
|
|
|
|
### 1. Synthetic Trigger Message (Plugin-Side)
|
|
|
|
After detecting NO_REPLY and advancing the turn, the plugin sends a **synthetic message** to the channel that triggers the next agent.
|
|
|
|
**Challenges:**
|
|
- The plugin SDK (`message_sent` hook) doesn't have an API to inject messages into agent sessions
|
|
- Sending a real Discord message (even invisible like zero-width space) creates noise and may confuse other agents
|
|
- The synthetic message wouldn't contain the original user's context
|
|
|
|
### 2. Deferred Evaluation (Don't Block in before_model_resolve)
|
|
|
|
Instead of blocking non-speakers at `before_model_resolve`, let all agents receive the message but inject a "you are not the current speaker, reply NO_REPLY" instruction. The current speaker gets a normal prompt.
|
|
|
|
After the current speaker responds with NO_REPLY, the plugin would need to **re-trigger** the next agent's session with the same message.
|
|
|
|
**Challenges:**
|
|
- All agents still consume tokens for the NO_REPLY evaluation
|
|
- Re-triggering a session with an already-processed message requires OpenClaw internal APIs
|
|
|
|
### 3. Queue + Replay (Plugin-Side State)
|
|
|
|
The plugin stores the original message when it arrives. After NO_REPLY, it replays the message by injecting it into the next speaker's session.
|
|
|
|
**Challenges:**
|
|
- Requires access to session injection API (not available in current plugin SDK)
|
|
- Managing the message queue adds complexity
|
|
|
|
### 4. Gateway-Level Support (OpenClaw Core Change)
|
|
|
|
Add a plugin hook return value like `{ defer: true }` in `before_model_resolve` that tells OpenClaw: "don't process this message yet, but keep it pending." When the turn advances, the plugin could call `api.retrigger(sessionKey)` to replay the pending message.
|
|
|
|
**Challenges:**
|
|
- Requires changes to OpenClaw core, not just the plugin
|
|
- Needs design discussion with the OpenClaw team
|
|
|
|
### 5. Bot-to-Bot Handoff via Discord Message
|
|
|
|
When current speaker NO_REPLYs, have **that bot** send a brief handoff message in the channel: e.g., "（轮到下一位）" or a reaction. This real Discord message triggers all agents, and the turn manager ensures only the next speaker responds.
|
|
|
|
**Challenges:**
|
|
- Adds visible noise to the channel (could use a convention like a specific emoji reaction)
|
|
- The no-reply'd bot can't send messages (it was silenced)
|
|
- Could use the discord-control-api to send as a different bot
|
|
|
|
### 6. Timer-Based Retry (Pragmatic)
|
|
|
|
After advancing the turn, set a short timer (e.g., 2-3 seconds). If no new message has arrived, send a minimal trigger. This could be an internal "nudge" if the SDK supports it.
|
|
|
|
**Challenges:**
|
|
- Timing is fragile
|
|
- Still needs a mechanism to trigger the next agent
|
|
|
|
## Recommendation
|
|
|
|
**Solution 5 (Bot-to-Bot Handoff)** is the most pragmatic with current constraints. The implementation would be:
|
|
|
|
1. In the `message_sent` hook, after detecting NO_REPLY and advancing the turn:
|
|
2. Use the discord-control-api to send a short message (e.g., `[轮转]` or a specific emoji) from the **next speaker's bot account** in the channel
|
|
3. This real Discord message triggers OpenClaw to route it to all agents
|
|
4. The turn manager allows only the (now-current) next speaker to respond
|
|
5. The next speaker sees the original conversation context in their session history and responds appropriately
|
|
|
|
**Downside:** Adds a visible "[轮转]" message. Could be mitigated by immediately deleting it after delivery, or using a reaction instead of a message.
|
|
|
|
## Open Questions
|
|
|
|
1. Does the OpenClaw plugin SDK support injecting messages into sessions?
|
|
2. Can plugins access the Discord client to send messages directly?
|
|
3. Would an OpenClaw core `defer`/`retrigger` mechanism be feasible?
|
|
4. Is visible channel noise acceptable for the handoff message?
|