Feedback Learning Loop

MultiModel Dev OS implements a feedback learning loop, converting developer edits and instruction overrides into reusable prompt rules.

---

1. Learning Flow

When an agent proposes changes and the developer modifies or rejects the code:

1. Directive Compilation: The developer logs the feedback comment:

   npx multimodel-dev-os feedback add "Do not use Tailwind CSS in components under src/components/legacy; use CSS modules." --type preference --tags styling

2. Summarization: The developer compiles the logged guidelines:

   npx multimodel-dev-os feedback summarize

3. Learnings Update: The learning rules are compiled and saved to .ai/intelligence/learning-rules.md.

---

2. Dynamic Context Ingestion

During subsequent task prompts:

1. Pattern Matching: The routing engine scans learning-rules.md for rules matching the current file targets.
2. Constraint Assembly: Matching rules are compiled and injected directly into the prompt context budget as strict formatting rules.
3. Instruction Drift Prevention: This guarantees that once a design preference is specified, developer agents immediately align to it without repeating mistakes.

---

Command Center Integration

The Repository Command Center dashboard counts the logged feedback items and tracks whether learning-rules.md is present or missing. Running mmdo status will advise you when feedback logs exist but have not yet been summarized into active rules.