迷你部落格

My token budget is starting to burn fast. Here are the token optimizations I've done to reduce the cost for OpenCode.

• Install RTK.
  • Reduce shell output. See here for the supported commands.
• Install Codegraph.
  • Remember to codegraph init new repository !
  • Make sure that the MCP server is live.
• Install OpenSlimedit.
  • The code source is very tiny. It basically compress built-in tool descriptions that are frequently used, and perform additional little optimisations here and there.
• Use cheaper models for subagents.

Excluded plugins

• Dynamic Context Pruning Plugin : Not particularily convinced that it would reduce cost. Especially when I read their section Impact on Prompt Caching that kinda confirm my skepticism :

  Trade-off: Pruning reduces context size but can increase cache misses. The cost balance depends on your conversation, compression frequency, and provider pricing.

  Cache misses were exactly a significant part of my tokens cost. No point to risk it only to figure out what "intelligently" managed conversation ever mean.

• OpenCode Snip : Used to reduce the output of shell commands. Most certainly redunding with RTK.

除了主要使用的LLM之外，有時會用到Deepseek的API。

上個月，該公司公佈了要改採尖離峰價，價格差達兩倍之多。尖峰時間，簡單而言就是北京上班時間（不包括中午休息及大陸地區假日）。

對東亞時區的使用者來說，Deepseek或許不再有低價的吸引力。對歐洲就絲毫無任何改變，因為離峰時間對應的是下午至午夜。

某位兄臺為此做了一面DeepSeek尖離峰時段價格時鐘，供大家參考。

這時代還有些大公司仍用著千篇一律的拖曳式拼圖Captcha。

VLM一下就能低成本寫出一個CV-based的通解^*，也不消仰賴諸如2Captcha的人類API了。

這麼多年了，Captcha依舊阻礙人類，卻讓程式通行無阻，何嘗不是一種逆圖靈測試?

亦即是說成本比放置Captcha的公司還低（零）。換言之，公司在花錢降低UX。