Release Notes: Adaptive Token Compressor#

Version 2026.3.0#

Release Date: November, 2026

Summary:

Adds an experimental third compressor type for conversation context, and supports a small reranker model in the tool compressor for faster tool selection.

New:

  • Experimental: ContextCompressor ("context" compressor type) optimizes long, tool-heavy conversation histories through turn-aware policies and selective compression, using tool-interaction boundaries to retain recent turns and process older content while preserving call-result associations. It replaces eligible oversized historical results with recovery-guided stubs, and can preserve reads of files later changed in the conversation. Remaining eligible content can be routed through Claw Compactor or Compress Server.

  • Small-model tool selection with Qwen/Qwen3-Reranker-0.6B (predictor_backend="reranker"): about 10x lower predictor latency than the default Qwen/Qwen3.6-35B-A3B configuration, at a lower tool-token saving

  • New ToolCompressor parameters top_k, tool_count_level, predictor_max_output_tokens, and predictor_max_input_tokens

  • Multi-key metrics aggregation in CompressionManager: declare metrics_keys when registering a compressor (for example, ["provider", "user"]) and pass metrics_values to compress() to aggregate metrics independently by each key’s values; retrieve results with snapshot(key=..., value=...) or snapshot_by_keys().

Upgrade Notes:

  • Avg*PerRequest metrics now divide by the unique requests that went through the metric’s own sources, instead of all requests seen by the CompressionManager. Requests that did not call any of the metric’s compressors no longer lower its average, and snapshot() and snapshot(source=...) return the same value for the same metric. Values change only when some requests skip some compressors. Without req_id from the metric’s sources, the set_per_anchor() call count is still the denominator.

  • CompressionManager.snapshot(source=...) now selects metrics by the compressor they track instead of by metric name: it returns the metrics registered with sources set to the given compressor’s registered name, whatever the metric names are. Metrics named <name>.<metric> that track only that compressor are returned as before. Metrics that track several compressors, and RequestCount metrics, are no longer returned even if their names start with <name>.; read them with snapshot().

Version 2026.2.0#

Release Date: September, 2026

Summary:

Initial release of the Adaptive Token Compressor — a pluggable compression library purpose-built for LLM agent systems. Using a single unified interface, it compresses system prompts, context, and tool schemas across an LLM agent pipeline to reduce token usage and improve inference efficiency.

Features:

  • Unified compressor API with two compression types: conversation messages (harness) and tool descriptions (tool)

  • Factory-based construction for drop-in integration as a plugin in other projects

  • LLMLingua-backed text compression for local Lingua Server backends (PyTorch/OpenVINO)

  • LLM-based tool selection through a configurable predictor endpoint

  • Hybrid rule-based and model-based compression to balance compression ratio and content fidelity

  • Configurable tool-injection placements to flexibly trade off token savings against prefix-cache hit rate

  • Per-compressor telemetry for tokens, savings, compression ratio, and latency, with cross-compressor aggregation through CompressionManager