Learning Rules and Feedback Loop

MultiModel Dev OS captures developer corrections and instruction overrides to compile reusable system rules that refine subsequent agent proposal generations.

---

1. Feedback Loop Commands

1. Log Feedback

To log a design preference or instruction override, run:

npx multimodel-dev-os feedback add "Prefer vanilla CSS or CSS Modules in legacy components" --type preference --tags styling --files "src/components/legacy/*"

This logs a structured JSON entry to .ai/intelligence/feedback-log.jsonl.

2. List Logged Feedback

To view all feedback entries in a formatted list:

npx multimodel-dev-os feedback list

3. Summarize Feedback

To compile the raw feedback log into instructions that agents read:

npx multimodel-dev-os feedback summarize

This aggregates the logs by type and writes .ai/intelligence/learning-rules.md.

---

2. Context Ingestion by Agents

Prior to generating proposals or refactoring components:

1. Model agents read .ai/intelligence/learning-rules.md as context.
2. The rules act as prompt constraints matching specific file target patterns.
3. This prevents agents from repeating previously corrected mistakes.