LLMSlim v0.4.0

LLMSlim v0.4.0 — Tool-aware context, without rewriting the contract

Release date: 2026-08-20

What changed

LLMSlim now has stable infrastructure for representing, inspecting,
canonicalizing, fingerprinting, verifying, and measuring supported tool
contracts without modifying the authoritative execution schema. The release
also includes experimental retrieval research, kept separate from stable APIs.

Stable tool-schema infrastructure

• ToolSchema provider normalization for supported MCP, OpenAI function,
  Anthropic, and generic tool definitions.
• Deterministic canonical JSON and complete-contract SHA-256 fingerprints.
• Exact-equivalence verification and safe catalog optimization.
• Raw-schema preservation: the caller/executor remains authoritative.

Experimental tool retrieval

RESEARCH ONLY — NOT ENABLED AUTOMATICALLY.

TF-IDF, BM25, optional intfloat/multilingual-e5-small dense retrieval,
BM25+dense RRF, selective exposure, and lazy hydration are model-context
planning experiments. They do not authorize or execute a tool. Low-confidence
policies fail open to the full catalog.

Install the optional local semantic support with:

pip install "llmslim[semantic]"

The normal pip install llmslim path stays lightweight. The semantic model is
pinned to revision 0e60b8d9d2166d80387f86e3b48ec9ced55f4d15, is loaded from
an explicit local cache only, and is not included in package artifacts.

Benchmark findings

Phase 4 measured 375 schemas across 18 catalogs. Lossless serialization
reduced the compact canonical baseline by 0 tokens (0.00%); the baseline was
already compact, so this is a valid result rather than a hidden failure.

Phase 4.5/4.6 retrieval results are research evidence, not marketing claims.
On the frozen corpus, BM25 had 98.25% all-required recall with 25.89%
selective coverage; dense and hybrid reached 98.54% recall but failed open
more often and avoided fewer tokens. Median tokens avoided was zero for every
strategy. The safety–selectivity frontier did not improve.

External ToolRet validation was attempted but not completed within the declared
CPU/resource budget. No ToolRet score is claimed.

Security boundaries

Authoritative schemas remain authoritative. Ranking is not authorization;
tool annotations and _meta are untrusted for retrieval representation where
excluded, remote references are not fetched automatically, and LLMSlim does
not cryptographically authenticate caller or provider provenance. It remains
defense in depth, not complete prompt-injection prevention.

Upgrade

Default compress() behavior is unchanged. Tool-contract APIs are available
from llmslim.tools; retrieval modules are explicitly experimental. Review
docs/tool-apis.md and SECURITY.md before
integrating tool workflows.

LLMSlim v0.3.1

Highlights

• Adds provenance-aware ContextRole handling so untrusted RAG, tool, and assistant content cannot gain protected priority solely from imperative wording.
• Adds caller-supplied rewrite providers and validated rewrite/hybrid workflows.
• Improves token counting, inline-code handling, CJK behavior, and compression selection correctness.
• Makes the benchmark runner report actual pytest outcomes.

Verification

• 432 tests passed with 0 failures
• 92.57% branch coverage
• Ruff passed
• Benchmark suite passed on 43 bundled samples across 8 dataset files

Notes

v0.3.1 is a Python release. npm, Rust, ONNX, and WASM runtimes are not shipped. Rewrite and hybrid modes require a caller-supplied provider.

See the repository release notes and docs/release/v0.3.1-release-report.md for details.

v0.3.0

llmslim v0.3.0 — Hybrid Prompt Compression & Semantic Optimization

v0.2.0

Release v0.2.0 — High-Performance Semantic Prompt Compression