The author of the open-source project Tura reports that using macro execution with Codex CLI 0.144.1 and GPT-5.6-sol significantly reduced interaction rounds and costs across 60 DeepSWE debugging and rewrite sessions.

  • Macro + backward reasoning achieved an 80.0% pass rate (48/60) using 2,017 rounds and $221.138.
  • Macro Direct achieved a 65.0% pass rate (39/60) with only 969 rounds and $99.620.
  • Compared to Codex CLI High, Macro Direct used 84.0% fewer rounds and 83.5% fewer observed tokens while maintaining a 5-point higher pass rate.
  • The macro approach defines deterministic workflows below the model boundary, returning combined evidence only when new reasoning decisions are required.

This method allows users to save budget on stronger investigation and verification by stopping macros whenever new evidence requires a decision, rather than blind batching.