Demo: Safe Improvement Loop

Who: Teams who want their AI agents to propose codebase improvements with human-in-the-loop safety. Time: ~2 minutes | Prerequisites: Node.js 18+, a MultiModel Dev OS workspace

---

Starting State

You have a project with MultiModel Dev OS initialized and at least one coding session completed. You've noticed patterns you want the system to learn.

---

Workflow

Step 1: Record developer feedback

npx multimodel-dev-os@latest feedback add "Always use TypeScript strict mode" --type preference
npx multimodel-dev-os@latest feedback add "Avoid Tailwind CSS in this project" --type override

What happens: Logs feedback entries to .ai/intelligence/feedback-log.jsonl (gitignored). Each entry is timestamped and typed.

Step 2: Compile feedback into learning rules

npx multimodel-dev-os@latest feedback summarize

What happens: Reads all feedback entries and compiles them into .ai/intelligence/learning-rules.md — a human-readable file that AI agents can reference. Gitignored.

Step 3: Propose an improvement

npx multimodel-dev-os@latest improve propose --title "Add TypeScript strict config"

What happens: Creates a structured proposal file under .ai/proposals/ with frontmatter, description, and optionally a deterministic operations JSON block.

Step 4: Review proposals

npx multimodel-dev-os@latest improve review
npx multimodel-dev-os@latest improve status

What happens: Lists all proposals and their statuses (draft, approved, applied, refused). Read-only.

Step 5: Validate safety gates

npx multimodel-dev-os@latest improve validate .ai/proposals/proposal-XXXX.md

What happens: Runs 12 safety gates including path boundary checks, protected path blocks, approval status, and operation type validation. Read-only.

Step 6: Preview changes

npx multimodel-dev-os@latest improve diff .ai/proposals/proposal-XXXX.md

What happens: Shows a unified diff of what the proposal would change. No files are modified.

Step 7: Apply the approved proposal

npx multimodel-dev-os@latest improve apply .ai/proposals/proposal-XXXX.md --approved

What happens: Executes the deterministic operations (create_file, append_line, replace_text). Records execution in .ai/proposals/apply-log.jsonl with SHA-256 hashes.

Step 8: Check the audit log

npx multimodel-dev-os@latest improve log

---

What Gets Created

File	Purpose	Gitignored?
.ai/intelligence/feedback-log.jsonl	Raw feedback entries	✅
.ai/intelligence/learning-rules.md	Compiled learning rules	✅
.ai/proposals/proposal-XXXX.md	Improvement proposal	❌ (reviewable)
.ai/proposals/apply-log.jsonl	Execution audit log	✅

---

Safety Notes

• Steps 1-6 are non-destructive — feedback logging, proposal drafting, review, and diff are all safe
• Step 7 modifies files but only through validated deterministic operations
• 12 safety gates must pass before any apply can execute
• Proposals require explicit --approved flag
• All applied changes are audited with pre/post SHA-256 file hashes
• Protected paths (.git/, .env, node_modules/, apply-log.jsonl) are blocked

---

Cleanup

# Remove feedback and proposals (gitignored files stay local)
rm -f .ai/intelligence/feedback-log.jsonl
rm -f .ai/intelligence/learning-rules.md
rm -rf .ai/proposals/

---

Next Steps

• Build memory index: npx multimodel-dev-os@latest memory build
• Run a workflow cycle: npx multimodel-dev-os@latest workflow run feedback-review
• Hand off to next agent: Multi-Agent Handoff Demo