Thinking tool

The thinking tool enables Kiro to show its reasoning process when working through complex problems, providing transparency into how conclusions are reached.

Overview

When enabled, the thinking tool allows Kiro to break down complex reasoning into visible steps, helping you understand the decision-making process and learn from the AI's approach to problem-solving.

Enabling thinking tool

kiro-cli settings chat.enableThinking true

Or use the experiment command:

/experiment
# Select "Thinking" from the list

How it works

When Kiro encounters a complex problem that requires multi-step reasoning, it will:

1. Show its thought process as it works through the problem
2. Break down the reasoning into logical steps
3. Explain intermediate conclusions before reaching the final answer
4. Make the decision-making transparent so you can follow along

When thinking appears

The thinking tool is automatically invoked by Kiro when:

• Complex problem-solving requires multiple logical steps
• Analyzing trade-offs between different approaches
• Planning multi-step implementations with dependencies
• Debugging complex issues that need systematic analysis
• Making architectural decisions with multiple considerations

Example usage

Without thinking tool

> What's the best way to implement caching for our API?

I recommend using Redis for caching your API responses...

With thinking tool

> What's the best way to implement caching for our API?

🧠 Thinking...

Let me think through the caching requirements:

1. First, I need to consider the type of data being cached
   - API responses suggest read-heavy workload
   - Need to handle cache invalidation

2. Evaluating caching solutions:
   - In-memory (fast but limited by RAM)
   - Redis (distributed, persistent, feature-rich)
   - Memcached (simple, fast, but less features)

3. Considering your use case:
   - API responses benefit from distributed cache
   - Need TTL support for automatic expiration
   - May need cache invalidation on updates

4. Redis appears best because:
   - Supports complex data structures
   - Built-in TTL and eviction policies
   - Can be distributed across nodes
   - Persistent if needed

Based on this analysis, I recommend using Redis for caching your API responses...

Benefits

For learning

• Understand reasoning patterns: See how complex problems are broken down
• Learn problem-solving approaches: Observe systematic analysis
• Identify gaps in logic: Spot where reasoning might be flawed

For debugging

• Trace decision paths: Understand why certain recommendations were made
• Verify assumptions: Check if the AI's assumptions match your context
• Catch errors early: Spot logical errors before implementation

For complex tasks

• Follow multi-step plans: See how complex tasks are decomposed
• Understand dependencies: Identify relationships between steps
• Validate approach: Confirm the reasoning aligns with your goals

Use cases

Architectural decisions

> Should we use microservices or a monolith for our new project?

🧠 Thinking...

Let me analyze this architectural decision:

1. Project characteristics to consider:
   - Team size and structure
   - Expected scale and growth
   - Deployment requirements
   - Development velocity needs

2. Microservices advantages:
   - Independent scaling
   - Technology flexibility
   - Isolated failures
   - But: increased complexity, operational overhead

3. Monolith advantages:
   - Simpler deployment
   - Easier debugging
   - Lower operational cost
   - But: scaling limitations, tight coupling

4. Decision factors:
   - Small team → Monolith likely better
   - Unclear requirements → Start with monolith
   - Need rapid iteration → Monolith faster initially
   - Can refactor to microservices later if needed

Based on this analysis...

Algorithm selection

> What sorting algorithm should I use for this data?

🧠 Thinking...

Let me think through the sorting requirements:

1. Data characteristics:
   - Size of dataset
   - Is data partially sorted?
   - Memory constraints
   - Stability requirements

2. Algorithm options:
   - Quick Sort: O(n log n) average, O(n²) worst
   - Merge Sort: O(n log n) guaranteed, stable
   - Heap Sort: O(n log n), in-place
   - Tim Sort: Optimized for real-world data

3. Trade-offs:
   - If data is partially sorted → Tim Sort excels
   - If memory is limited → Heap Sort (in-place)
   - If stability matters → Merge Sort or Tim Sort
   - For general case → Quick Sort often fastest

4. Recommendation depends on...

Debugging complex issues

> My application is slow but I don't know why

🧠 Thinking...

Let me systematically analyze potential performance issues:

1. Common performance bottlenecks:
   - Database queries (N+1 problem, missing indexes)
   - Network latency (API calls, external services)
   - Memory issues (leaks, excessive allocation)
   - CPU-bound operations (complex calculations)

2. Diagnostic approach:
   - First, profile to identify hotspots
   - Check database query patterns
   - Monitor network requests
   - Analyze memory usage patterns

3. Most likely culprits based on typical patterns:
   - Database queries often account for 60-80% of slowness
   - Should start there before optimizing code

4. Recommended investigation order...

Configuration

Enable/disable

# Enable
kiro-cli settings chat.enableThinking true

# Disable
kiro-cli settings chat.enableThinking false

# Check status
kiro-cli settings chat.enableThinking

Limitations

Performance impact

• Slower responses: Thinking adds time to generate responses
• More tokens used: Thinking process consumes additional tokens
• Longer output: Responses are more verbose

When not to use

• Simple questions: Thinking overhead not needed for straightforward queries
• Quick iterations: May slow down rapid back-and-forth
• Well-defined tasks: When you just need the answer, not the reasoning

Best practices

When to enable

• Learning new concepts: Understand the reasoning behind recommendations
• Complex decisions: Need to validate the decision-making process
• Debugging: Want to see systematic problem analysis
• Code reviews: Understand why certain approaches are suggested

When to disable

• Simple tasks: Straightforward questions don't need reasoning
• Speed priority: Need quick responses without explanation
• Familiar territory: Already understand the reasoning patterns

Workflow integration

1. Enable for complex tasks: Turn on when starting difficult problems
2. Learn from reasoning: Observe how problems are broken down
3. Disable for execution: Turn off when implementing known solutions
4. Re-enable for review: Turn back on when validating approaches

Troubleshooting

Thinking not appearing

1. Verify it's enabled:

   bash

   kiro-cli settings chat.enableThinking
   

2. Try a complex question: Simple questions may not trigger thinking

3. Restart chat session: Changes may require new session

Too much thinking

If thinking output is overwhelming:

1. Disable for simpler tasks:

   bash

   kiro-cli settings chat.enableThinking false
   

2. Ask for concise answers: Request brief responses explicitly

3. Use for specific questions: Enable only when needed

Related features

• Experimental Features
• Tangent Mode - Explore reasoning without affecting main conversation
• TODO Lists - Break down complex tasks

Next steps

• Enable other experimental features
• Configure settings
• Learn about custom agents