TLDR Tech / AI
Model-Harness-Fit
Bustamante dissects Codex CLI, Claude Code, and GitHub Copilot CLI to show that frontier labs post-train models against specific harnesses, baking tool names, schemas, citation tags, memory rituals, and system prompt structures into the weights. Terminal-Bench 2.0 data backs the thesis: Claude Opus 4.6 scored 79.8% with ForgeCode versus 75.3% with Capy, and Cursor jumped from "Top 30 to Top 5" by changing only the harness, while OpenAI models default to patch-based file edits and Anthropic models to string replacement, with mismatches costing reasoning tokens.
Read the full story