opencode-sop-engine puts skills on rails

This feels like the right answer for teams that are tired of “best effort” agent behavior and want process they can actually enforce. Hard gating in `tool.execute.before` is stronger than prompt-only guardrails because invalid calls can be blocked before they happen. Requiring `submit_step_result` with real `evidence.callIds` makes “looks done” outputs much harder to fake. The history-pruning angle matters: long-context degradation is a real failure mode in agent workflows, especially after retries and detours. The tradeoff is operational overhead; somebody has to author and maintain the workflow profiles, so this is best for teams that already want standardized SOPs. The project reads more like agent governance infrastructure than a flashy end-user app, which is exactly why it may age well.