/**
* Subagent Summarize Pattern
*
* Instructions for using Claude Code's Task tool to achieve RLM-style
* context isolation when retrieving Thoughtbox sessions.
*/
export const SUBAGENT_SUMMARIZE_CONTENT = `# Subagent Summarize Pattern
Use Claude Code's Task tool to retrieve and summarize Thoughtbox sessions **without polluting your conversation context**.
---
## The Problem
When you call Thoughtbox MCP tools directly, the full response appears in your conversation:
\`\`\`
You: [call session.get]
Thoughtbox: [returns 50 thoughts, 8000 tokens]
↓
All 8000 tokens are now in YOUR context
\`\`\`
## The Solution
Spawn a sub-agent to do the retrieval. Only its summary returns to you:
\`\`\`
You: [spawn sub-agent via Task tool]
Sub-agent: [calls Thoughtbox, retrieves 50 thoughts]
Sub-agent: [summarizes internally]
Sub-agent: [returns 200-token summary]
↓
Only 200 tokens in YOUR context
\`\`\`
**10-40x context reduction** while preserving key information.
---
## How to Use
### Summarize a Session
\`\`\`json
{
"tool": "Task",
"subagent_type": "general-purpose",
"description": "Summarize Thoughtbox session",
"prompt": "Retrieve and summarize Thoughtbox session.\\n\\n1. Call mcp__thoughtbox__init with operation 'get_state'\\n2. Call mcp__thoughtbox__thoughtbox_cipher\\n3. Call mcp__thoughtbox__session with operation 'get' and sessionId '<SESSION_ID>'\\n4. Summarize the key insights in 3-5 sentences\\n\\nReturn ONLY your summary. Do not include raw thought content."
}
\`\`\`
### Search Across Sessions
\`\`\`json
{
"tool": "Task",
"subagent_type": "general-purpose",
"description": "Search Thoughtbox sessions",
"prompt": "Search Thoughtbox for information about <TOPIC>.\\n\\n1. Call mcp__thoughtbox__init with operation 'get_state'\\n2. Call mcp__thoughtbox__thoughtbox_cipher\\n3. Call mcp__thoughtbox__session with operation 'list'\\n4. For relevant sessions, call operation 'get' to retrieve thoughts\\n5. Extract only information related to <TOPIC>\\n\\nReturn a concise summary of findings. Do not include raw thought content."
}
\`\`\`
### Synthesize Multiple Sessions
\`\`\`json
{
"tool": "Task",
"subagent_type": "general-purpose",
"description": "Synthesize Thoughtbox sessions",
"prompt": "Synthesize conclusions across multiple Thoughtbox sessions.\\n\\n1. Initialize Thoughtbox (init → cipher)\\n2. List sessions with tags matching '<TAG>'\\n3. Retrieve each relevant session\\n4. Identify common themes, contradictions, and key conclusions\\n5. Synthesize into a coherent summary\\n\\nReturn only your synthesis. Do not include raw thoughts."
}
\`\`\`
---
## Why This Works
Claude Code's Task tool spawns an isolated sub-agent:
- Sub-agent has its own conversation context
- Sub-agent's MCP calls stay in ITS context
- Only the sub-agent's final output returns to parent
- Parent never sees the full MCP responses
This is **RLM-style context isolation** using existing primitives.
---
## Template Variables
Replace these in the prompts above:
| Variable | Description |
|----------|-------------|
| \`<SESSION_ID>\` | UUID of session to retrieve |
| \`<TOPIC>\` | Search topic or keyword |
| \`<TAG>\` | Session tag to filter by |
---
## Best Practices
1. **Be specific about output format** - Tell sub-agent exactly what to return
2. **Forbid raw content** - Explicitly say "do not include raw thought content"
3. **Set token budget** - "Summarize in 3-5 sentences" or "max 200 words"
4. **Use haiku for simple tasks** - Add \`"model": "haiku"\` for faster/cheaper sub-agents
---
## Example: Full Invocation
\`\`\`typescript
// In your Claude Code conversation:
const result = await Task({
subagent_type: "general-purpose",
model: "haiku", // Cheaper for summarization
description: "Summarize auth session",
prompt: \`
Retrieve and summarize Thoughtbox session about authentication.
Steps:
1. mcp__thoughtbox__init({ operation: "get_state" })
2. mcp__thoughtbox__thoughtbox_cipher()
3. mcp__thoughtbox__session({ operation: "get", args: { sessionId: "abc-123" }})
4. Extract key conclusions about authentication decisions
Return a 3-sentence summary. No raw thoughts.
\`
});
// result contains ONLY the summary, not the full session
\`\`\`
---
## Experimental Results
Tested 2026-01-16:
| Approach | Context Cost | Info Quality |
|----------|--------------|--------------|
| Direct MCP call | ~800 tokens | Full raw data |
| Sub-agent pattern | ~80 tokens | Summarized insights |
**10x reduction confirmed.**
---
## Alternative: Via Gateway Tool
If tool list doesn't refresh (common with streaming HTTP clients), use the gateway:
\`\`\`json
{
"tool": "Task",
"subagent_type": "general-purpose",
"description": "Summarize Thoughtbox session via gateway",
"prompt": "Retrieve and summarize session via gateway tool.\\n\\n1. mcp__thoughtbox__thoughtbox_gateway({ operation: 'start_new', args: { task: 'retrieve' } })\\n2. mcp__thoughtbox__thoughtbox_gateway({ operation: 'cipher' })\\n3. mcp__thoughtbox__thoughtbox_gateway({ operation: 'session', args: { operation: 'get', args: { sessionId: '<SESSION_ID>' } } })\\n4. Summarize key insights (3-5 sentences)\\n\\nReturn ONLY your summary. Do not include raw thought content."
}
\`\`\`
### When to Use Gateway
Use the gateway approach when:
- Client doesn't refresh tool lists properly
- Using streaming HTTP transport
- Other Thoughtbox tools appear unavailable
The gateway tool is always enabled at Stage 0 and routes to all other handlers internally.
---
## See Also
- \`thoughtbox_cipher\` - Compress thought content (complementary approach)
- \`session.get\` - Direct retrieval (when you need full content)
- \`thoughtbox_gateway\` - Always-available router for all operations
- RLM paper: https://arxiv.org/abs/2512.24601
`;
