Summary

Add the MLflow tracking implementation as a concrete example in the AI provider agnosticism principle documentation.

Context

We just implemented truly provider-agnostic MLflow tracking in PR #1328 that demonstrates the principle perfectly by:

• Avoiding N×M complexity (N providers × M patterns)
• Focusing on extracting actual commands rather than provider-specific formatting
• Using a single parser (parse_session.py) for all AI assistants

What to Document

In knowledge/principles/ai-provider-agnosticism.md

Add a new section showing MLflow tracking as a real-world example:

## Implementation Example: MLflow Session Tracking

The MLflow tracking system demonstrates provider agnosticism by extracting actual commands from transcripts rather than maintaining provider-specific patterns:

**Anti-pattern (N×M complexity):**
- Different regex patterns for each provider
- Provider detection logic
- Maintenance burden grows with each new AI assistant

**Correct pattern (provider-agnostic):**
- Single parser looks for actual `git`, `gh`, and bash commands
- No provider detection needed
- Works automatically with any AI assistant

See: `tracking/parse_session.py` - Extracts real commands regardless of AI formatting

Benefits

• Shows a concrete implementation of the principle
• Demonstrates how to avoid the N×M problem
• Provides a reference for future provider-agnostic implementations

Related

• PR feat: make MLflow tracking AI provider agnostic #1328: MLflow provider-agnostic implementation
• Issue feat: make MLflow tracking AI provider agnostic (support Codex, etc.) #1326: Original MLflow tracking feature request