Multi-AI Critique with Memory

Ask an AI to review your code and it tells you the code is good. That is not a review — it is a mirror.

Pipeline Pro is built the other way round: two reviewers on different models critique every change against a protocol written to find defects, not to approve work. The verdict is binding enough to make you stop and argue when it rejects you.

Then it remembers. Every critique is embedded at 768 dimensions and stored. Before the next review, it searches that history by cosine similarity — no cap, relevance decides what comes back. A pipeline that has reviewed your codebase for months knows your patterns. Catching and fixing the spaghetti code your amnesiac coding AI has quietly been creating.

Three protocols:

• Design: Is this the right thing to build? (before writing code)
• Code: Is this change correct? What else is wrong nearby? (after pushing)
• Verify: Did the approved change land correctly? (after merging)

What you get: critique_v3.py worker, two reviewer protocols on different AI models, GitHub Actions workflows, schema_migration.sql, preflight and memory backfill scripts, SETUP.md written to hand to your AI assistant (~30 minutes to install).

What you bring: GitHub account, Supabase free tier or your own Postgres, OpenAI key, optionally Gemini (~10x cheaper). A few cents per review. Nothing hosted, no telemetry, no licence server.