blas0

Here is something pretty cool that I encountered during a personal project session...

Intent verification for AI-pilled devs

We've all been there. You ask Claude for a login form with Zod validation, it ships something that looks right, you move on. Three files later you realize...no Zod anywhere. The vibes were immaculate but the implementation was mid.

Built a hook-based daemon that shadows Claude Code sessions. Captures your original intent, watches every Write/Edit, then runs a review agent that actually checks if the code matches what you asked for. Catches yolo implementations before they compound into tech debt.

Zero config. Just hooks. Findings get injected back into context so Claude fixes its own mistakes.

CAM addresses a fundamental LLM limitation; statelessness. Each session starts fresh. Claude Hooks became the architectural driver to solve this.

The mechanism: Claude Hooks intercept operations (UserPromptSubmit, PostToolUse, SessionEnd, etc.) and feed data into CAM—code changes, tool use patterns, research, prompts, plans. CAM embeds everything, builds relationships, and injects relevant context before Claude processes requests. Allowing contextually aware planning, tool use, researching, diagnosing, debugging, implementations, etc.

The result: an infinitely updated, persisted codebase map + documentation source. CAM reads/writes/iterates/revises/queries/validates research and tool use across sessions, compounding context awareness over time.

This pattern generalizes beyond Claude. Any LLM integration with hooks can gain persistent memory this way.

Open source MVP. Uses Gemini embeddings, SciPy clustering, knowledge graphs.