In :doc:`V2 <v2>`, each iteration issues one MMA and waits for it to complete (via tcgen05.commit + mbarrier.wait) before issuing the next. The tensor core MMA pipeline sees only one operation at a ...

dsh-llm-replay lets snapshot tests run the real agent without an API key by replaying model streams from recorded Session JSONL fixtures. Each parent and subagent session receives its recorded script ...