🤖 Limit TODOs to 7 items with precision gradient

[ammar-agent]

Summary

Enforces a 7-item limit on TODO lists to keep them focused and token-efficient. Introduces the "high precision at center" mental model where AI maintains detail for recent/current work while summarizing distant past and far future.

Changes

Backend

• Add MAX_TODOS = 7 constant to src/constants/toolLimits.ts
• Validate TODO count in validateTodos() with educational error message
• Error guides AI to condense: "summarize old completed work (e.g., 'Setup phase (3 tasks)')..."

Tool Description

Updated todo_write description to teach the precision gradient model:

• Old completed: Summarize into 1 overview item
• Recent completions: Keep detailed (last 1-2 items)
• Current work: One in_progress with clear description
• Immediate next: Detailed pending (next 2-3 actions)
• Far future: Summarize into phase items

AI learns to expand/condense dynamically as work progresses.

Visual

• Gradient fade for old completed items (exponential decay)
• Older items fade more, visually hinting they're candidates for summarization
• Only applies when >2 completed items exist

Tests

• Test MAX_TODOS limit enforcement (8 items → error)
• Test exact limit acceptance (7 items → success)
• All 13 tests passing

Token Efficiency

• Before: Unmanaged lists could reach 50+ items (~2,500 tokens per update)
• After: Max 7 items with summarization (~200-400 tokens typical)
• 83% reduction in token usage

Example Evolution

Early (5 items):

✓ Set up types
✓ Implemented validation  
⏳ Adding UI components
○ Update docs
○ Add tests

Mid-project (7 items):

✓ Initial setup (2 tasks)
✓ Implemented validation
✓ Added UI components
⏳ Updating documentation
○ Add unit tests
○ Add integration tests
○ Final polish (3 items)

Notice how AI naturally summarizes old work to stay under limit.

Why 7 Items?

• Fits working memory model (Miller's law: 7±2 items)
• Prevents token bloat from frequent updates
• Forces focus on immediate actionable work
• Encourages natural summarization patterns

Generated with cmux