llmslim Files

LLMSlim Core 0.11.0 puts Claude Code and LangChain agents on a lean context. Long test logs, build output and search results are shortened the first time the model sees them, then the same bytes are sent on every later turn, so prompt caching and Claude's thinking keep working.

:::bash
pip install --upgrade "llmslim[proxy]"
llmslim wrap claude

Highlights

• llmslim wrap claude runs Claude Code through a local proxy. Shell, search and web output is shortened once, keeping errors and the lines related to the command, and replayed byte for byte afterwards. File reads and edits are never touched. On Claude Haiku 5.5 with adaptive thinking, a 10,545-token pytest log went out as 156 tokens, cache reads grew every turn, and 62,334 input tokens were not sent across seven requests. The same request without the replay was rejected with HTTP 400. Guide: docs/PROXY.md (github.com).
• LLMSlimToolOutputMiddleware does the same for LangChain 1.x create_agent (pip install "llmslim[langchain-agents]").
• JSON results keep every item, one compact line each. An 80-issue GitHub list went from 12,247 to 3,108 tokens with every issue listed.
• PostgresStore lets several servers share one memory (pip install "llmslim[postgres]").
• The TypeScript client is on npm: npm install llmslim, with Vercel AI SDK tools in llmslim/ai-sdk.

The full list is in the [CHANGELOG](https://github.com/Thanatos9404/llmslim/blob/main/CHANGELOG.md#0110.