Compact Context: a local five-file starting map for Codex

[mithatb]

I built Compact Context, an MIT-licensed, local repository router for Codex.

The idea is deliberately small: before a qualifying coding turn, it ranks local repository metadata and gives Codex up to five likely files. Codex still has normal access to the full repository and reads the authoritative source before editing.

It uses Python's standard library, makes no additional model calls, does not upload source code, and abstains when the match is weak.

Measured result

On a frozen 68-task benchmark across 10 public repositories and four ecosystems:

• starting repository-map payload: 1,402.1 → 230.4 mean tokens
• reduction: 83.6%
• target file among five starting suggestions: 48/68 → 50/68

Important limitation: this measures the starting-map payload, not total Codex tokens or billing, and the retrieval difference is not statistically established as a quality improvement.

Try it

curl -fsSL 'https://compact-context.met.chatgpt.site/install.sh?v=4' | bash

Repository: https://github.com/mithatb/compact-context
Benchmark and row-level data are included in the repository.

I'd especially value feedback from people using Codex on larger Python, JavaScript/TypeScript, Go, or Rust repositories:

• Did the five-file shortlist help you start?
• Where did it miss or add noise?
• Which repository structures break the ranking?

Failures and reproducible counterexamples are very welcome.