Repository navigation
Force agent to use CMM? #34
Description
Is there any way to force the agent to always use MCC for search as the default, or is that the "wrong" way to use it?
Or have I misinterpreted how it's used by agents?
For example, here's the heuristic it came back with when I asked how it decided when to switch:
My heuristic:
Use grep_search when:
• I know (or can guess) the exact string/symbol — function name, variable, import path
• I want to find where something is used across files
• The codebase is small enough that raw text search is fast and precise
• I need current file contents (grep always reflects the live files)
Use codebase-memory-mcp when:
• The question is conceptual — "how does X work", "what calls Y", "what's the architecture of Z"
• I need a call graph or dependency trace across many files
• grep would require 5+ sequential searches to piece together a picture
• The repo has been indexed and the graph is fresh
The practical tell: If I can write the search string in 5 seconds, I use grep. If I'd need to follow a chain of "grep → read result → grep again → read result" more than 2-3 times, the MCP graph traversal is cheaper.
Activity
sonicviz commentedon Mar 15, 2026
I'm just asking if I misinterpreted how it's supposed to work via agents like copilot.
Is the above correct, or not?
It's not exactly clear in the docs how and when you can trigger it or if you have any control in agent mode unless you invoke directly via terminal?
DeusData commentedon Mar 19, 2026
Fixed in 159378a and 442cb95 — shipping with v0.5.0.
The install command now enforces MCP tool preference through multiple mechanisms per agent:
The heuristic you described (grep for exact strings, MCP for conceptual/multi-hop queries) is exactly right — our instructions explicitly tell agents when to fall back to built-in search. The hooks are advisory (exit 0), not blocking.
Thanks for the detailed analysis @sonicviz — it directly shaped the instruction content.
If agents still default to grep too often after v0.5.0, please reopen or file a new issue.
sonicviz commentedon Mar 20, 2026
Thanks. Does this also install for CoPilot in VSCode?
Another issue, not sure if it needs a report, is when you update it installs a bunch of integrations which are not installed, causing issues with the agent trying to start noni-runnable MCP servers as the base app is not installed.
I don't have cursor, windsurg, or zed but it installed them. I only use CoPilot (as main) and claude/gemini cli as alternates.
CoPilot lets you choose any model easily, and the agentic harness is as good as any of the others.
Plus, it's subsidized, best cost-effective deal you can get for models like opus.