Codeset lifts Codex with repo context

This is a credible argument for making code agents more repo-native instead of bolting on a separate retrieval stack. The gains are modest on SWE-Bench Pro but strong enough to suggest the biggest wins come from grounding agents in local conventions, not generic reasoning.

• –The larger lift on codeset-gym-python is meaningful, but it is also the benchmark the team controls, so the public verifiers matter more than the raw percentage
• –The smaller SWE-Bench Pro delta suggests repo context helps most on tasks where local history, file relationships, and test habits matter
• –Static in-repo files are operationally simple: no vector DB, no runtime RAG service, no extra infra for the agent to call
• –The tradeoff is staleness; if the repo changes quickly, extracted context will need refreshes to stay useful
• –For teams already using coding agents, this looks like a low-friction upgrade worth testing before building heavier retrieval plumbing