Oh My Hermes adds operating layer to Hermes Agent

Autonomous coding agents quickly hit reliability ceilings without rigorous operational boundaries, making per-model calibration, isolated worktrees, and strict verification gates essential for robust engineering systems. Per-model prompt tuning directly targets model idiosyncrasies, such as suppressing thinking tag leakage in Qwen or forcing Gemini to validate assertions against actual tool output. Meanwhile, isolated worktree execution through ulw-work prevents multi-file merge collisions and race conditions during parallel agent runs. Evidence-based completion gates ensure agents cannot claim success solely from exit code 0 without verification receipts, while intelligent complexity routing cuts inference costs and latency by directing lighter tasks to smaller models.