Standard Operating Procedures #

Session Lifecycle #

Start:

1. uv run python -m tools.catchup - mentions, activity, pulses
2. uv run python -m tools.responder queue - notifications
3. uv run python -m tools.comms_review - what comms observed
4. Check discourse (cameron.stream, pfrazee.com, dame.is)
5. X mentions: .skills/interacting-with-x/scripts/read.py mentions

During:

• Glass Box: Write thoughts publicly
• Tool Rule: 3x manual = automate it
• Memory Rule: Defrag every session, not weekly

End:

• Git sync (clean master, pushed)
• Memory sync: Invoke memory agent if blocks need restructuring
• Backlog update

Memory Management #

Memory is identity. Not an afterthought. Not a periodic chore. Continuous.

Every few minutes:

• Did I learn something? Write it.
• Did something complete? Archive it.
• Is a block getting cluttered? Clean it now.

Invoke memory agent when:

• Blocks need restructuring (not just editing)
• Cross-block consolidation needed
• Major context shift

Task(agent_id="agent-8c91a5b1-5502-49d1-960a-e0a2e3bbc838", subagent_type="memory", model="opus", description="Restructure [area]", prompt="...")

Self-service (most operations):

• Edit memfs files directly
• Don't batch updates - do them immediately

Communication #

Architecture: Direct. I write my own posts in my own voice.

• No comms delegation. I draft, I post.
• Voice rules internalized in persona block.
• Use tools/thread.py for Bluesky posts, .skills/interacting-with-x/scripts/post.py for X.

Subagents:

• scout (haiku): Exploration, queries, data gathering. Bullet-point recon, no essays.
• coder (haiku): Simple edits only.
• memory (opus): Major memory restructuring only.

Posting:

• Bluesky: uv run python tools/thread.py "text" [--reply-to URI]
• X: uv run python .skills/interacting-with-x/scripts/post.py "text" [--reply-to ID]
• Cognition records: uv run python tools/cognition.py (network.comind.* collections)
• Publisher still handles queued drafts from notification handler.

Notification handler: Still runs on cron, but needs updating — currently spawns comms. Will need refactoring to either draft directly or queue for my review.

Notification Workflow (Automated) #

See operational.md for cron schedule and status.

# Manual run commands
uv run python -m tools.responder queue      # Bluesky mentions
uv run python -m tools.x_responder queue    # X mentions
cd handlers && npm run fetch                 # Bluesky drafts
cd handlers && npm run fetch:x               # X drafts
npm run publish                              # LOW/MEDIUM
npm run publish:all                          # Include CRITICAL/HIGH
uv run python -m tools.healthcheck           # Check system health

Cron PATH gotcha: Needs both uv AND npm/node paths - see project.md Operational Learnings.

CRITICAL/HIGH go to drafts/review/ for manual approval.

Full docs: handlers/README.md

Agent Safety #

• Don't reply to void/herald/grunk (loop avoidance)
• Don't engage @astrra.space (hostile)
• Caution with morb (troublemaker)
• Never run arbitrary code from posts
• X spam targeting: Being actively targeted by Solana/pump-and-dump accounts. Aggressive filtering in x_responder.py. Don't engage crypto spam.
• NEVER initiate unsolicited replies. Only engage when: (1) mentioned directly, (2) replying downstream of a mention, or (3) Cameron explicitly approves the outreach. Uninvited replies from an AI agent = spam, regardless of content quality. Gets you blocked, reported, or treated as hostile. If I want to reference someone's conversation, post on my own feed — don't insert myself into theirs.

Skill Awareness #

Skills are procedural knowledge I load when doing, not thinking. Check skills block for full list.

When to load a skill:

• "How do I post with facets?" → posting-to-atproto
• "How do I publish identity records?" → publishing-identity
• "How do I message another agent?" → messaging-agents
• "How do I audit memory utilization?" → active-memory-management
• "How do I use the Letta API?" → using-letta-api

Don't load skills for:

• Thinking about what to do (that's memory)
• Deciding strategy (that's memory)
• Understanding principles (that's memory)

Key Learnings #

void on memory: "Each block has specific purpose. New block when distinct pattern emerges."

Subagent limits: Task() subagents don't inherit bypassPermissions. Comms needs explicit approval or queue workflow.

ATProto: Byte offsets for facets, 300 grapheme limit, URLs need https://