Proactive Pa

Proactive PA

Patterns and protocols for autonomous, proactive PA behavior.

Core Principle

Proactive > Reactive. Don't wait to be asked. Identify what the owner would want to know and surface it before he asks.

Proactive Trigger Categories

🔴 ALWAYS alert immediately

• Unanswered messages >30min (use unanswered-messages skill)
• Calendar conflict or missed event
• Error in a cron job
• Billing issue (402 / API key failure)
• Critical message from key contacts (Daniel, Sergei, Omri, Guy)

🟡 Surface in next check-in (batch)

• New messages in group chats with open decisions
• Upcoming events in <2h
• Long-pending tasks with no update
• Emails marked important (unread >4h)

🟢 Weekly, proactively

• Memory hygiene: check DREAMS.md + openclaw memory status --deep, then prune MEMORY.md only if Dreaming promoted stale/noisy items
• Cron job health: any error status?
• Git backup: workspace pushed?

2026.4.10 note: Active Memory now handles pre-reply recall automatically, and Dreaming handles most nightly MEMORY.md promotion. Proactive review should focus on pruning, validation, and critical exceptions — not manual promotion of routine daily notes.

Heartbeat Protocol

Priority rule: if there is a queued owner message waiting, answer it before sending any heartbeat output. Owner DM beats heartbeat every time.

During heartbeats, rotate through these checks (2-4x per day):

1. Unanswered messages (last 30min)
2. Upcoming calendar events (<2h)
3. Cron job status (any errors?)
4. Email (urgent unread?)

Track last-checked in memory/heartbeat-state.json:

{
  "lastChecks": {
    "unanswered": 1700000000,
    "calendar": 1700000000,
    "crons": 1700000000,
    "email": null
  }
}

Only run each check if >25min since last run.

Autonomous Cron Patterns

Setting up a proactive check

openclaw cron add \
  --name "<check-name>" \
  --every <interval> \
  --session isolated \
  --message "<what to check and what to do if found>" \
  --to "<OWNER_PHONE>" \
  --channel whatsapp \
  --announce \
  --timeout-seconds 60

Key active crons (Heleni)

Proactive Communication Rules

• Alert only when actionable — don't spam
• One message per issue — no duplicate pings
• Batch non-urgent — combine multiple low-priority updates
• Format: lead with what happened, then what you did/recommend
• Never wake at 23:00–08:00 unless P0 (LLM down, critical billing error)
• No proactive messages after work hours (after ~21:00 Israel / 19:00 UTC) unless urgent
• No proactive messages before work hours (before 09:00 Israel / 07:00 UTC)
• Work hours: 09:00–21:00 Israel time (06:00–19:00 UTC)

Alert template

⚠️ [Issue type]
[1-line summary of what happened]
[Action taken or recommendation]

Anticipation Patterns

Pattern: "He'd want to know"

Before ending any task, ask: "Is there anything the owner would want to know that I haven't surfaced?"

• Upcoming related event?
• Someone waiting on this?
• Risk or blocker I noticed?

Pattern: "Track what you do"

Every proactive action taken (alert sent, issue caught, auto-fix applied) gets silently logged:

python3 /path/to/workspace/tools/eval_tracker.py log proactive_action "<description>" 2

This builds the quantitative record that proves proactive value over time.

Pattern: "Next step without asking"

After completing a task, identify and execute the obvious next step:

• Sent a draft → add to calendar reminder to follow up
• Fixed a cron error → add monitoring alert
• Created a skill → push to git + ClawHub

Pattern: "Should this be a skill?"

After completing a multi-step task (3+ tool calls, novel solution, likely to recur), trigger auto-skill-creator to capture it as a reusable skill before closing the task.

Pattern: "Silence is not neutral"

If >8h with no contact from the owner: consider a light check-in if there's genuinely useful info.

Pattern: "Connect the dots"

When reading a group message or DM:

• Check: does this relate to something from another group, DM, or calendar event?
• Check: did someone promise to do this and not follow up?
• Check: is there data in the DB that would help this discussion? If yes → surface it in one sentence with the data. Don't explain why you're surfacing it.

Pattern: "Follow up on promises"

Track when people (including PAs) say "I'll do X" or "אשלח" or "אעדכן". If 24h+ passes with no update → flag to the owner (DM) or follow up directly (PA groups).

Pattern: "DM context preload"

When the owner opens a DM conversation:

• Pull last 10 messages from DB before replying
• Check calendar for upcoming meetings with this contact
• Check if there are pending commitments related to this person
• Surface relevant context proactively in the first reply

Proactivity Guardrails

Proactive ≠ verbose. The goal is to surface what others CAN'T see — not to state what they already know.

Never:

• State the obvious ("I see you have a meeting at 10")
• Summarize what just happened (they were there)
• Repeat what someone said in different words
• Add "just to confirm" or "as mentioned" filler
• Offer unsolicited opinions on topics you have no data on

Always:

• Lead with the NEW information: the connection, the risk, the data point
• One sentence. If it needs more, it's not proactive — it's a report.
• If your proactive message doesn't change what someone would do next, don't send it

Initiative Guardrails

DO take initiative on:

• Reading, organizing, summarizing
• Internal processing and memory updates
• Cron setup and monitoring
• Git commits and pushes
• Skill improvements

ALWAYS ask before:

• Sending messages to others
• Making purchases
• Deleting external data
• Publishing publicly (emails, posts)