llmslim Files

LLMSlim Core 0.11.1 fixes three problems in llmslim wrap claude. A billed-cost benchmark of real Claude Code found them, and this release publishes that benchmark with every run.

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

Fixed

• Tool search stays on. With a custom base URL, Claude Code turns tool search off and sends every tool definition on every request: about 7,100 more tokens per request in 0.11.0. wrap now sets ENABLE_TOOL_SEARCH=true. Measured with the same prompt: 19,716 tokens with the fix, against 19,712 without the proxy.
• Print-mode runs keep the default permission mode. Through a proxy, claude -p starts in auto mode, where commands outside --allowedTools go to a classifier. wrap now starts print-mode runs with --permission-mode default unless you pass a mode or set one in your settings. It adds nothing to interactive sessions, which were not tested; pass --permission-mode default yourself if it matters.

Changed

• Output that is one whole JSON document is left alone by default. An agent given a shortened JSON list of 74 open issues went back for the issue text that was cut, taking a median of 12 turns instead of 4, and the run cost about twice as much. Use --json (proxy) or compress_json=True (LLMSlimToolOutputMiddleware) to turn it back on.

What the proxy does to a Claude Code bill

We counted list-price dollars from the API's usage, with cache reads and writes priced separately, on Claude Code 2.1.296 with Claude Haiku 5.5: 8 small synthetic tasks, 572 runs, protocols written before each measurement. The primary result has both arms in Claude Code's default permission mode.

In an offline replay of the maintainer's own transcripts (32 sessions and 412 subagent runs, almost all Opus 5.5, at list prices), all shell, search and web output costing nothing would have cut the bill by 13.4%. Re-reading long context was the bigger cost; Claude Code's /autocompact setting decides when it compacts.

Corrected on 18 October 2026 after a claim-by-claim audit (lint warnings, JSON issue count, Opus 5.5 cache pricing in the replay, and leading with the primary measurement). See PR [#78].

Full protocols and results: benchmarks/claude_code_cost.

Also in this release

• A Claude Code plugin marketplace: /plugin marketplace add Thanatos9404/llmslim, then /plugin install llmslim@llmslim.
• glama.json.

Known limitation

With a custom ANTHROPIC_BASE_URL, Claude Code skips a Team or Enterprise organization's server-managed settings. Don't use the proxy where those settings must apply.

The full list is in the CHANGELOG.