v0.11.2

Five opt-in capabilities that strengthen the core promise: measurable, query-aware, recoverable, structure-preserving compression. Every addition is backward-compatible. Default output is byte-identical to v0.11.1 unless a feature is explicitly enabled.

Added

• Session token-savings telemetry. The registry now derives savings from each tool's originalTokens/optimizedTokens and feeds a process-scoped session accumulator, fixing the long-standing tokensSaved = 0 hardcode. Set DISTILL_SAVINGS_STATS=1 to append a compact savings line (per call + session total) to tool output.
• JSON/YAML structural skeletons in smart_file_read. Config files (package.json, tsconfig, k8s manifests, OpenAPI specs) now return a key-outline skeleton instead of a full-file dump: native JSON.parse for a typed tree, an indent-based walk for YAML, no new dependency. Explicit mode: "full" still returns the raw file.
• Query-aware semantic compression. auto_optimize accepts an optional task param. When set, the semantic compressor lifts segments covering more task terms in its selection ranking, so task-relevant context survives compression. With no task, behavior is unchanged.
• Opt-in reversibility (ctx.restore). A bounded in-memory origin store (process-scoped, no disk state) keeps pre-compression originals keyed by a content-derived handle. With DISTILL_RETRIEVE=1, auto_optimize emits the handle and code_execute exposes ctx.restore(handle) to recover the original when a lossy pass dropped something the agent needs.
• Constraint preservation in conversation compaction. ctx.conversation.compress now extracts binding constraints (must, never, do not, always, required) and pins them verbatim in a [Preserved constraints] block, so lossy summarization can no longer silently drop a rule the downstream model still has to obey.

Fixed

• ToolResult.tokensSaved is now populated instead of always 0, re-enabling the verbose saved: diagnostic line.

Full changelog: https://github.com/ArthurDEV44/distill/blob/main/CHANGELOG.md