Description
Background
Modern code assistants like Claude 4 or GPT-4 are increasingly agentic—they don't just generate one-off responses, but plan multi-step actions (e.g. “refactor this file, then update the test cases”). However, current systems only record the final code output, leaving out the reasoning process and intermediate intentions.
Problem
We lose critical context:
Why was this change made?
What alternatives were considered?
What was the agent’s overall plan or subgoals?
This makes auditing, debugging, or retracing AI-generated code hard.
Proposal
Introduce MCP (Model Context Protocol) support to:
Record the agent’s plan, step-by-step (via tool_call traces or structured plans).
Track corresponding code diffs for each plan step.
Allow replay or inspection of an agent’s thought process alongside its edits.
Benefits
Improve transparency and reproducibility of AI-driven coding.
Enable better debugging, rollback, and education workflows.
Pave the way for more collaborative agent workflows (“show me what the agent was going to do before it did it”).
Example Flow
Agent plans: "Step 1: refactor X. Step 2: update Y."
Each step triggers a snap with diff + plan metadata.
Final commit includes not just code, but the full agentic trace.
Next Steps
MCP should expose a way to bind structured plans to snapshots.
Allow storing plan+diff pairs as atomic steps in Memov.
Optional: support replaying the trace in the UI.