1. Is this an official DeepSeek site?
No. It is an independent notes site written from real runs by a user. Nothing here is an official statement, and no page uses DeepSeek branding. See关于页。
2. Why does a plugin I installed make the app hang with no output?
Most likely a missing client.inject package. Measured: a plugin naming six client packages the host lacks produced seventy seconds of zero output, a live process, and no error. There is no version-gate refusal to look for. See插件核验表。
3. Why did the version gate not stop me installing an incompatible plugin?
Because it does not exist. Measured on 0.1.2-alpha.5: a plugin requiring >=0.2.0-rc.2 installed fine, composed fine in --dump-config, and was refused by nothing. Of the 87 first-party packages, zero declare engines.dsh. See版本门那条。
4. My token numbers are much lower than yours. Why?
Probably reading only some of the three figures. Measured: on one session, reading the last usage per turn gave 28,827 while summing per step gave 126,551. Also, inputTokens is cache-miss input only; cache hits are separate. See实测量法。
5. Is a cache hit cheaper even when total tokens go up?
Yes. Measured: run two used 51 percent more total tokens than run one at a sixth of the cost, because 47,872 of its tokens were cache hits at $0.003 per million against 16,210 cache-miss tokens at $0.15. See缓存命中那条。
6. Which number should I optimise — tokens or money?
Money, and within money the output line. Measured across six tasks, output accounted for 59 to 69 percent of cost because output is priced 200 times a cache-hit input token. Cutting output and step count saves more than cutting input. See成本基线。
7. Can I trust the English pages for exact wording?
For commands and numbers, yes. For prose, no. Every English page is a generated translation with no native review. Where a detail is ambiguous, the Chinese page governs. See译校状态。