context-aware-delegation（上下文感知委托）

Context-Aware Delegation

(aka "SmartBeat")

Problem: Isolated sessions (cron jobs, sub-agents) can't see your main session conversation history. They're cheap (use Haiku) but blind to context.

Solution: Use sessions_history to give isolated sessions full awareness of what happened in your main chat — at a fraction of the cost of running everything in main session.

Quick Start

Morning Report Example

You want a daily report that includes "what we accomplished last night" — but running that in main session with Sonnet costs ~$0.30/report. Using an isolated session with Haiku costs ~$0.03, but can't see conversation history.

Solution: Isolated session queries main session history first.

// Inside your cron payload.message:
"1. Query main session history: sessions_history('agent:main:telegram:direct:{userId}', limit=50)
2. Read memory files: memory/YYYY-MM-DD.md
3. Fetch weather for Austin 78721
4. Generate report combining:
   - Recent conversation highlights
   - Memory file summaries
   - Current conditions
5. Send via Telegram + email"

Cost: ~$0.03 with Haiku (10x cheaper than Sonnet main session) Context: Full awareness of overnight work

Pattern Overview

1. Identify Main Session Key

# List sessions to find main
sessions_list(limit=10)
# Typical main session key format:
# agent:main:telegram:direct:{userId}
# agent:main:main

2. Query History from Isolated Session

// In cron job, sub-agent, or event handler:
sessions_history({
  sessionKey: "agent:main:telegram:direct:8264585335",
  limit: 50  // Last 50 messages
})

Returns conversation history even though you're in an isolated session.

3. Use Context + Execute Task

Your isolated session now has:

• ✅ Conversation history (what was discussed)
• ✅ Memory files (persistent notes)
• ✅ Cheap model (Haiku)
• ✅ Full tool access

Use Cases

Cron Jobs with Context

Morning reports:

Schedule: 8 AM daily
Model: Haiku (~$0.03/run)
Task: Read overnight work, check email, send summary
Context: Last 50 messages from main session

End-of-day summaries:

Schedule: 9 PM daily
Model: Haiku
Task: What got done today? What's pending?
Context: Today's full conversation

Periodic check-ins:

Schedule: Every 2 hours (9 AM - 9 PM)
Model: Haiku
Task: Anything urgent in email/calendar?
Context: Recent discussion about priorities

Sub-Agent Delegation

Background builds:

sessions_spawn({
  task: "Build the AREF product page based on our discussion",
  model: "haiku",
  // In the task prompt:
  // "First, query main session history to see our conversation about AREF requirements..."
})

Research tasks:

sessions_spawn({
  task: "Research Unreal Engine integration patterns. Reference our earlier discussion about AREF goals.",
  model: "haiku"
})

Event-Driven Handlers

Webhook arrives → isolated session handles it:

// Webhook payload triggers isolated session
// Session logic:
"1. Query main session to see: what did J and I agree about this client?
2. Process webhook based on that context
3. Take action or notify"

Cost Comparison

Savings: ~10x cheaper than main session, with same context awareness.

Implementation Tips

Finding Your Main Session Key

sessions_list({ kinds: ["main"], limit: 5 })
// Or:
sessions_list({ limit: 10 })
// Look for: agent:main:telegram:direct:{yourUserId}

How Much History?

• 10 messages: Just recent context (~2KB)
• 50 messages: Last few hours of work (~10KB)
• 100 messages: Full day or multi-session context (~20KB)

Start with 50, adjust based on needs.

Combining History + Memory

Best results come from:

1. Sessions history: Recent interactive work
2. Memory files: Persistent decisions/notes

"1. sessions_history(limit=30) → what we discussed today
2. read memory/2026-02-13.md → decisions logged
3. Combine both sources for complete picture"

Morning Report Recipe

Complete example for daily morning report:

Cron Job Setup:

{
  schedule: { kind: "cron", expr: "0 8 * * *", tz: "America/Chicago" },
  sessionTarget: "isolated",
  payload: {
    kind: "agentTurn",
    model: "haiku",
    message: `Generate morning report:

1. Query main session: sessions_history('agent:main:telegram:direct:8264585335', limit=50)
2. Read yesterday's memory: memory/YYYY-MM-DD.md
3. Get weather: Austin 78721
4. Check email (gog or himalaya)
5. Check calendar events for today

Report format:
📍 WEATHER: [conditions]
🌙 OVERNIGHT: [from session history - what we worked on]
📝 PERSISTENT NOTES: [from memory file]
📧 EMAIL: [urgent only]
📅 CALENDAR: [today's events]
🔗 DASHBOARD: [mission control link]

Send to Telegram using message tool.

Note: Email delivery from isolated sessions requires SMTP credentials or is better handled via main session heartbeats for reliability.`
  },
  delivery: { mode: "announce", to: "8264585335", channel: "telegram" }
}

Cost: ~$0.03/report (~$1/month) Context: Full overnight work awareness Timing: Exact (8 AM every day)

Limitations

History truncation:

• sessions_history returns limited content (typically last N messages)
• Very long messages may be truncated
• For deep archives, rely on memory files

Main session must exist:

• If main session is brand new (no messages), history is empty
• Isolated sessions can't create main session history, only read it

Not real-time:

• History reflects state when queried
• If main session is actively running, very latest messages might not appear immediately

Best Practices

1. Write good memory summaries Even with session history access, persistent memory files are gold. Don't rely solely on conversation history.

2. Query only what you need limit=10 for quick context, limit=50 for substantial work, limit=100 for deep dives.

3. Chain tools effectively

sessions_history → memory_get → web_search → message

Context first, then action.

4. Use Haiku for delegation, Sonnet for decisions

• Isolated background work: Haiku
• Interactive problem-solving: Sonnet
• Morning reports/summaries: Haiku
• Architecture discussions: Sonnet

Troubleshooting

"Empty session history"

• Check session key is correct: sessions_list()
• Main session might be new (no messages yet)
• Use limit parameter

"Content truncated"

• Reduce limit (fewer messages = more complete content)
• Rely on memory files for archival data

"Isolated session can't send messages"

• Use message tool, not sessions_send
• Ensure delivery.mode is set in cron config OR use message tool directly

Related Patterns

• Heartbeats: Main session periodic checks (full context, main model)
• Sub-agents: Long-running background tasks
• Cron jobs: Scheduled isolated work
• Memory files: Persistent cross-session storage

Credits

Discovered by RGBA Research during OpenClaw optimization work. Published to ClawHub as open pattern for the community.

Contact: https://rgbaresearch.com License: MIT (free to use, adapt, share)