0xWhiteMage

Deterministic, local context compaction with query-aware tool-result extraction and recovery. Disclosure — Subclasses Hermes's built-in ContextCompressor and overrides its private prune/demote methods (checks their shape at startup and falls back to stock compaction when they change). Unless you set them yourself, turns on proactive tool-result pruning at compression.proactive_prune_tokens=48000 and lowers proactive_prune_min_result_chars to 2000. Adds a model-callable tameru_expand tool that returns pruned tool output from memory or from this session's archive; complex regex greps run in a short-lived local Python child process. No network calls; the render log is off by default.