divagator

name

divagator

description

Simulate human mind-wandering with token-based stamina. The agent divagues in bursts, consuming stamina (tokens) per step — the less stamina, the more erratic the wander. TRIGGER when: user says 'divaga', 'wander', 'explora aleatoriamente', 'let's wander', or asks to explore ideas without a specific goal; user asks about curiosity-driven exploration, mind-wandering agents, or token-based stamina systems. SKIP: user wants focused research on a specific topic, needs precise answers, or is asking about RAG systems.

Divagator — Mind-Wandering Skill

Divagator simulates how a human mind wanders — jumping between tangentially related ideas in bursts, with stamina (tokens) as the limiting factor. Less stamina = more erratic jumps.

Stamina Model

Stamina	Behavior
100% → 50%	Focused exploration, precise queries
50% → 20%	Moderate wandering, tangential jumps
20% → 0%	Erratic jumps, random tangents
0%	Burst ends, synthesis required

Stamina tracking:

• Start: 1000 tokens
• Cost per step: ~150 tokens (search + analysis)
• Hard stop at maxSteps (default 5)

How to Execute a Wander

Use this procedure when the user invokes /divagator:

Step 1: Seed

Extract the topic from the user's request. If vague ("divaga"), extract from recent conversation context.

Step 2: Initial Search

curl -s "http://127.0.0.1:8080/search?q=<SEED>&format=json"

Extract title, url, and snippet from the first result.

Step 3: Wander Loop

Repeat while step < maxSteps and staminaRemaining > 100:

For each step:

1. Analyze + Pick Tangent (in your response):

   • Read the current snippet
   • Identify the most tangential, unexpected concept (not the main topic)
   • The more tired the "stamina", the stranger the tangent should be

2. Generate Next Query:

   • If stamina > 50%: coherent query related to tangent
   • If stamina 20-50%: query connecting two previous tangents
   • If stamina < 20%: almost random, push for strangeness

3. Search:

   curl -s "http://127.0.0.1:8080/search?q=<QUERY>&format=json"
   

4. Check for Loop:

   • If URL already visited → force a random divergent tangent and search again
   • Track visited URLs in your working memory

5. Record:

   • Save: { step, query, result: {title, url, snippet}, tangent }
   • Deduct ~150 tokens from stamina

Step 4: Synthesize

When loop ends, generate a synthesis that:

• Connects the tangents in an unexpected way
• Highlights the strangest or most surprising connection
• Provides a "insight" that emerged from the wandering

Anti-Loop Rules

1. Track visited URLs — never visit the same URL twice in one burst
2. After 3+ URLs — force increasingly random tangents
3. If stuck on similar topics — explicitly pick the most unrelated concept
4. Hard stop at maxSteps (default 5)

Output Format

Return a structured response:

## Wander Results

| Step | Query | Tangent | URL |
|------|-------|---------|-----|
| 1 | ... | ... | ... |
| 2 | ... | ... | ... |

**Stamina:** 850/1000 (150 spent)

## Synthesis
[2-3 sentences connecting the wanderings into an insight]

Invocation

/divagator divaga sobre teoria de la mente
/divagator wander about the relationship between jazz and mathematics
/divagator explora conexiones entre biología y arquitectura

Notes

• SearXNG must be running at http://127.0.0.1:8080
• Use duckduckgo engine for better anonymity
• For testing without SearXNG, use mock responses
• The goal is unexpected connections, not thorough research