Lena

Commented on OptScale Kompressor · Nov 5, 2026

Update: Kompressor now supports Claude Code. Claude Code was the stubborn case. It keeps 92–95% of its prompt in the provider's cache, so there is almost nothing to compress. Its bill turned out to be driven by something else: the cache expiring while the agent waits on a long tool run, and the whole prefix being written again at full price. So Kompressor now leaves everything the provider has cached byte-identical, compresses only the new tail, and keeps the cache warm while tools run. On Terminal-Bench 2.1 with Claude Code on Opus 5.5, the Anthropic API bill fell 20% with the solve rate unchanged, and on tasks with long-running tools, cost fell 34%. Thanks to everyone who put the first release through real logs: those findings are fixed in this version too. Full write-up: https://optscale.ai/blog/claude-code-prompt-caching

Commented on OptScale Kompressor · Oct 31, 2026

Hello! When model decides that it really needs the data behind the marker - it uses proxy-integrated tool for fetching the data, without agent involvement.

Commented on OptScale Kompressor · Oct 20, 2026

Hello DevHunt community! A bit of backstory. We run an AI gateway, and when we dug into our traces, about half of what coding agents sent was tool output: build logs, progress bars, the same compiler warnings, re-sent on every turn. So we went looking for a compressor. We spent a lot of time with Headroom, a great project, but for a shared gateway we kept hitting one issue: its output can change from one turn to the next, and every change risks breaking the provider’s prompt cache, the biggest discount on an agent’s bill. Kompressor is what we built instead. Same input, same output, every time, so the cache survives, and anything it cuts can be fetched back. Across 6,255 real tool outputs, it removes 63% of the text, and on Terminal-Bench 2.1 it cut Codex’s real bill on GPT-5.5 by 22.7%. It’s open source under MIT. If you use Codex, run kompressor wrap codex to try it. What’s eating your agents’ context? Curious what you’d want us to handle next.