Context pre-filtering as a token optimization layer — benchmark data

[nicolalessi]

Hey @affaan-m, love the work on ECC. The token optimization and memory persistence sections address a real problem.
I've been working on a complementary approach: instead of optimizing at the prompt/session level (which ECC does well), pre-filtering what goes into the context window at the codebase level before the agent even starts.
Ran a controlled benchmark on FastAPI (~800 files):

Tool calls: 23 → 2.3 per task (-90%)
Cost: $0.78 → $0.33 per task (-58%)
Output tokens: 504 → 189 (-63%)

The approach uses tree-sitter AST parsing + dependency graph + SQLite, served via MCP. Everything local.
Curious if you've explored integrating a context pre-filtering layer into ECC's architecture. The two approaches (ECC for agent behavior + context engine for input optimization) could stack well.
Benchmark data: vexp.dev/benchmark
FastAPI writeup: reddit.com/r/ClaudeCode/comments/1rjra2w

[affaan-m]

Maintainer note: thanks for sharing the benchmark data. We are using this to inform ECC 2.0 context-filtering work and will follow up with our own measurements and implementation notes.