skyscribe

Software engineer at Hangzhou

[skyscribe @skyscribe]

@syhily 及时发后续

[skyscribe @skyscribe]

@fitchmultz this tool is nice, but i couldn't use it with kimi-k2.6 because of tool schema exceeds allowed size, is it possible to compress the json? I know this is a moonshot model limitation

[skyscribe @skyscribe]

@arunoda @akay_jpeg @antirez still over think and repeat itself for many occasions, quite annoying problem if you leave it to perform autonomous tedious work like fix broken ci

[skyscribe @skyscribe]

@akay_jpeg @antirez still it is quite difficult to subscribe to GLM token plans, either sold out or busy

[skyscribe @skyscribe]

发现qoder的套路真多，同样的CLI，海外版和cn版本的模型选择，300 credits对应的token数量，都有极大差异，看得出来大家不订阅保持观望是非常明智的选择。写个脚本统计token都得写两套，实际结果也叫人超级惊讶，细想又有那么一丝合理性在里面，不过不是用户角度而言

[skyscribe @skyscribe]

@nekocode_cn @chenchengpro 看你是否跑subagent，是否检查主工作区被意外污染。我用过的所有模型，都有一定概率拍到主工作区拉屎，导致workflow检查到污染暂停。尤其是你用大量的fan out+join merger的时候，中间某个subagent犯傻后面执行不下要求你来人工清理，那就损失很多时间，甚至无法恢复

[skyscribe @skyscribe]

@nekocode_cn @chenchengpro worktree容易不稳定污染，还是多目录好，我也切换多目录了

[skyscribe @skyscribe]

@amehochan 那个strReplace tool才是坑，好多次卡死了

[skyscribe @skyscribe]

@fitchmultz Windows with wsl2, which one to select

[skyscribe @skyscribe]

@Leechael @wey_gu 已经确认是pi subagent的问题，固定到前一个版本解决了。看起来昨天它们的更新导致单个tool的schema过大

[skyscribe @skyscribe]

@Phoenixyin13 @J88240S 老板正迷信大跃进，把软件工程师全部干掉呢。跨越四五个层级的决策大老板，看到可以砍掉三分之一到一半的人力，别提多振奋了，天天拉数据比较谁的AI账单最少，下达指挥系统鞭策各中底层管理岗

[skyscribe @skyscribe]

@Musiclover2000A @royxy 我觉得Deepseek写代码还是明显更啰嗦，出错更多的。价格嘛，如果是Deepseek v4 PRO，API直接付费肯定比Kimi 199套餐贵出来两三倍

[skyscribe @skyscribe]

@DangD51083018 @royxy 实际token统计上看，199算是量比较大的了。和其它友商的比较起来，又快又好，就是模型思维链太啰嗦，thinking比例大，国模通病。

[skyscribe @skyscribe]

@Leechael @wey_gu 一开始怀疑是task plane问题，因为它注册的tool特别多。后来卸载了task plane还不行，但是其它GLM5.1和GPT5.5都没有问题。我估计pi，subagent，Kimi privider都更新到最新应该100%复现。这阵子不在电脑前，不过半个小时前也反馈给pi subagent了

[skyscribe @skyscribe]

@Leechael @wey_gu 晚点回去查下，我一直都是不锁定版本，直接update跟最新的。不过前面出门之前查了一下，大概率是pi subagent的问题

[skyscribe @skyscribe]

@Leechael @wey_gu 今天早上一更新，pi里面选择Kimi coding直接报错了，提示tool schema超出限制，是subagent问题还是kimi provider膨胀了

[skyscribe @skyscribe]

@nicopreme after upgrade, i get problems with moonshot kimi model, it is complaining for schema exceeds maximum allowed size, any workaround?

[skyscribe @skyscribe]

@bimajatmiko_ @F2aldi Indeed subagents are heavy unfortunately, i also worked with task plane, but suffered a lot to remedy its internal states when things are not going well.

[skyscribe @skyscribe]

@evxnche @F2aldi same case https://t.co/oprazWkqow

[skyscribe @skyscribe]

@bimajatmiko_ @F2aldi The subagent extension is essential, the big thing is you can tune it on your needs by natural language talking to let pi install for you, even write and test tiny extensions for yourself