DSH + DeepSeek V4.1 Flash: should I reduce context window / reasoning?

I’m using DeepSeek V4.1 Flash inside DSH for fairly autonomous coding sessions. One recent ~1 hour session on a medium/large Flutter + backend repo did this: ~773 API requests ~198M total input tokens 197.5M were cache hits only ~818K cache misses ~408K output tokens ~99% cache hit ~313 tok/s Max reasoning cost was ~$0.96 during off-peak agent wrote ~5.5K lines, hit 2 CI failures, fixed them, and got CI green The weird part is that I previously used DeepSeek V4 Pro through Cline for 8–13 hours/day and would usually burn only ~$2/day. Now V4.1 Flash can burn ~$1 in roughly an hour despite being much cheaper per token. It looks like DSH is retaining a huge context/cache and repeatedly sending it across hundreds of tool calls. Average cached context for that session was roughly 255K tokens/request . For people experienced with DSH: Should I cap the context window? If yes, what would you recommend — 128K, 160K, 200K, 272K, etc.? Also, would you run High reasoning normally and Max only for difficult architecture/debugging , or keep Max enabled for autonomous coding? Main priority is code quality + agent reliability , but I don’t want to waste hundreds of millions of cached tokens if they’re no longer helping.