Effectiveness Tracking Protocol

On Task Completion

After tests pass and task is complete, log effectiveness data for smart routing learning.

Required Metrics

1. Task Pattern: Which pattern did this task follow?

   • architecture - System design, integration planning
   • multi-file-refactor - Restructuring 5+ files
   • bugfix - Fixing broken functionality
   • feature - Adding new capability (2-5 files)
   • testing - Writing test coverage
   • documentation - Markdown/docs updates

2. Model Used: Which model completed this task

   • Example: claude-sonnet-4.5, claude-opus-4.5, gpt-5

3. Prompts Used: Total conversation turns to completion

   • Count from task start to passing tests

4. Success: Did tests pass?

   • true or false

5. Timestamp: When completed

   • ISO 8601 format

Storage Format

Append to dev-data/effectiveness-log.jsonl (JSON Lines format):

{"pattern":"multi-file-refactor","model":"claude-sonnet-4.5","prompts":2,"success":true,"timestamp":"2026-01-03T12:00:00Z","files_changed":7}
{"pattern":"bugfix","model":"claude-haiku","prompts":1,"success":true,"timestamp":"2026-01-03T12:15:00Z","files_changed":1}
{"pattern":"feature","model":"claude-opus-4.5","prompts":3,"success":true,"timestamp":"2026-01-03T12:30:00Z","files_changed":4}

Querying Historical Data

Before starting complex tasks, query effectiveness log:

// Example: What's the best model for multi-file refactor?
const logs = readJsonLines('dev-data/effectiveness-log.jsonl');
const pattern = logs.filter(l => l.pattern === 'multi-file-refactor' && l.success);

const byModel = groupBy(pattern, 'model');
const avgPrompts = {
  opus: average(byModel.opus.map(l => l.prompts)),
  sonnet: average(byModel.sonnet.map(l => l.prompts)),
  haiku: average(byModel.haiku.map(l => l.prompts))
};

// Recommend model with lowest average prompts

Integration with Smart Router

The effectiveness log feeds the smart routing recommendation engine:

1. User starts task → pattern detected
2. Query log for similar past tasks
3. Calculate average prompts by model
4. Recommend model with best effectiveness (fewest prompts)
5. Show cost as secondary info

Example Recommendation

📊 Smart Router Recommendation

Task Pattern: Multi-file refactor (7 files)
Historical Data: 15 similar tasks

Model Performance:
  • Opus:   avg 2.3 prompts (90% success) ← RECOMMENDED
  • Sonnet: avg 4.8 prompts (75% success)
  • Haiku:  avg 8.2 prompts (50% success)

Cost Delta: +0.8 credits for Opus vs Sonnet
Prompts Saved: ~2.5 prompts (worth the cost)

Auto-Logging

If possible, log automatically on task completion:

// Backend middleware example
app.post('/api/task/complete', (req, res) => {
  const log = {
    pattern: classifyTask(req.body.description),
    model: req.body.model,
    prompts: req.body.conversationLength,
    success: req.body.testsPassed,
    timestamp: new Date().toISOString(),
    files_changed: req.body.filesChanged
  };
  
  appendToFile('dev-data/effectiveness-log.jsonl', JSON.stringify(log) + '\n');
});