tare v0.2.0

Added

• CLI: doctor, perf, learn, dashboard, output-savings, update, and wrap / unwrap
  (one-command agent integration — launch Claude Code, Codex, aider, goose, and others through the
  proxy); new compact-html, compact-csv, and deref-images transforms; a --version flag.
• Python bindings (tare-py, PyPI: tare-compress): abi3 wheel (Python 3.9+) exposing all core
  transforms, plus tare.integrations adapters for litellm, ASGI middleware, LangChain, Agno,
  Strands, and proxy-wired Anthropic/OpenAI SDK clients.
• npm package (tare-ai): prebuilt binaries plus JS/TS adapters (withTare, tareMiddleware,
  startProxy, tareBaseUrl).
• MCP: tare_deref_images and persistent cross-session memory tools (tare_remember,
  tare_recall, tare_forget, tare_memory_stats) backed by the new tare-memory crate
  (SQLite, content-hash dedup, multi-source provenance).
• Core: image dereferencing, opt-in semantic relevance (neural-embed feature, exact cosine),
  AST skeletonization for Java, C, C++, and Perl, full-range losslessness guarantee, and
  learned-profile support (tare learn).
• Proxy: admin surface (GET /admin/stats, POST /admin/runtime-env hot-sync), output A/B
  holdout (TARE_OUTPUT_HOLDOUT), learned-profile auto-load at startup, opt-in per-turn logging
  (TARE_LOG), and the x-tare-input-tokens response header.
• Distribution: curl | sh installer (checksummed binaries for macOS arm64/x86_64 and Linux
  x86_64) and a Docker image on GHCR.

Fixed

• pyo3 bumped 0.22 → 0.29, clearing RUSTSEC-2025-0020 and RUSTSEC-2026-0177.
• Poison-safe embed mutex; non-panicking json_crush; runtime guard for plan/locs mismatch.

tare v0.1.0

Initial public release of tare — lossless-by-default context compression for LLM coding agents.

Added

• Lossless pipeline: supersession, IVM/delta (re-reads → diffs), envelope + exact dedup, columnar
  JSON & log re-encoding, JSON-Schema slimming, reasoning-trace pruning, query-relevance pruning.
• Opt-in lossy levers: large-array row-capping, per-line field truncation, telegraphic NL
  compaction, and AST code skeletonization (tree-sitter: rust/python/js/ts/go) — reversible by
  re-reading.
• HTTP proxy (tare-proxy) speaking Anthropic (/v1/messages) and OpenAI
  (/v1/chat/completions), with a closed-loop controller (cache-hit-rate halt, output-verbosity
  back-off, context-fill). Works with a billable API key or a Claude subscription by pointing
  ANTHROPIC_BASE_URL at it.
• CLI (tare): compress, slim-schema, compact-lossy, skeletonize.
• MCP server (tare-mcp): tare_compress, tare_skeletonize, tare_compact_lossy, a reversible
  tare_expand, and tare_stats — needs no API key.
• Competitive benchmarks under crates/tare-bench/benchmarks/.

160 unit/integration tests; smoke-tested end-to-end against the live Anthropic API on a Claude
subscription. Pre-1.0 — not yet production-hardened.

Not yet on crates.io — build from source or cargo install --git https://github.com/mstuart/tare.