[KongchangAI]
Company

Atlarix

A development team for AI coding agents that builds an AI coding assistant with Ask and Build modes, with attention to inference optimizations such as prompt caching.

Timeline (last 90 days)

Oct 18

The Atlarix team's Build mode merged a stable header carrying a cache breakpoint with request text that changes every turn into the same block, causing the entire cache (about 19,000 tokens) to be rewritten for every message.

Unverified50%
Oct 18

By comparing two consecutive real request bodies in the same conversation, the team found that the first difference appeared at character 22,966 of the 24,000-character system block (containing the user request's <user_context>).

Unverified50%
Oct 18

The fix was to split the Build prompt into a cached portion containing no per-message changing content and a separately sent per-turn tail block, without changing the prompt text itself.

Unverified50%
Oct 18

Before the fix, the cache reads for messages 2 and 3 were 0, with cache writes of about 19,600 tokens per message.

Unverified50%
Oct 18

After the fix, for the fixed portion of each message, the cost on Claude is about 15 times cheaper than before the fix.

Unverified50%
Oct 18

This cache fix was released with Atlarix version 15.9.

Unverified50%

All Facts (6)

Source Articles