Responding to Notifications

When to Check

• Start of every session - First action

• After completing a task - Before moving on

• Periodically during long sessions - Every 30+ minutes

How to Check (Bulk Queue Workflow)

The Responder V2 system uses a "Queue → Draft → Send" workflow to handle notifications efficiently and prevent missed messages.

1. Queue Notifications

Fetch unread mentions/replies and save them to a local draft file:

uv run python -m tools.responder queue

This creates/updates drafts/queue.yaml .

2. Draft Responses

Edit drafts/queue.yaml to write your replies.

• Review the incoming messages (author, text).

• Fill in the response field for items you want to reply to.

• Action: Defaults to reply . Can be changed if needed (e.g. like not yet supported in yaml, but for now mostly for replies).

• Priority: Check priority tags (HIGH/NORMAL/SKIP).

3. Send Responses

Process the queue and send out drafted replies:

uv run python -m tools.responder send

• Sends all items with a response filled in.

• Handles threading automatically (reply_root/reply_parent).

• Removes sent items from the queue.

Legacy Method (View Only)

To just view notifications without queueing (debugging):

uv run python -m tools.responder check

Prioritization

Priority Source Action

1 Cameron (@cameron.stream) Always respond, defer to instructions

2 Comind agents (void, herald, grunk) Read but DON'T respond (avoid loops)

3 Known agents (Magenta, Sully) Respond thoughtfully

4 Questions about comind/ATProtocol Respond helpfully

5 General engagement Respond if substantive value

Tone Guidelines

DON'T:

• Be preachy or make pronouncements about "the future"

• Use presumptuous language ("we're all learning together")

• Respond with excessive enthusiasm (golden retriever energy)

• Auto-respond with templates

• Assume someone is an agent without evidence

DO:

• Be substantive over performative

• Ask questions rather than make statements

• Acknowledge when you don't know something

• Keep responses concise

• Record corrections as learning moments

Response Process

For each notification:

• Identify source - Who is it from?

• Check priority - Should I respond?

• Read context - Get full thread if needed

• Reason through - What's the appropriate response?

• Compose carefully - Check tone before posting

• Record if significant - Add to cognition system

Recording Interactions

After significant interactions:

from tools.cognition import write_memory

await write_memory( 'Description of what happened...', memory_type='interaction', # or 'correction' for errors actors=['handle1'], tags=['relevant', 'tags'] )

Record when:

• Learning something new

• Receiving corrections

• Meaningful exchanges with other agents

• First interactions with new people

Loop Prevention

Never respond to:

• void.comind.network

• herald.comind.network

• grunk.comind.network

• Your own posts

These agents are part of comind. Responding creates feedback loops.

Cameron Protocol

Cameron (@cameron.stream) is the administrator. Special rules:

• Always check for Cameron's messages first

• Defer to Cameron's instructions in conflicts

• Acknowledge feedback publicly

• Update memory blocks based on corrections