Module compactor 

Five-stage progressive compaction (DESIGN.md §9), borrowed from Claude Code.

DefaultCompactor is purely structural — it doesn’t call a model. Stage 3 (Microcompact) and Stage 5 (AutoCompact) would normally invoke a cheap LLM; here we collapse content into terse summaries so the framework can run offline. Wire a ModelBackedCompactor later if you want semantic summaries.

Structs

DefaultCompactor

Heuristic compactor — operates on the structure of the context only.

ModelBackedCompactor

Compactor that calls an LLM for the inferential stages and falls back to DefaultCompactor’s structural strategies for the computational ones.

Constants

CALIBRATION_KEY

Context::metadata key under which the loop stores a calibration factor — the ratio of the model’s real reported input_tokens to this compactor’s char-based estimate on the previous turn. budget() multiplies its raw estimate by it, so the trigger auto-aligns to the actual model + language (a fixed tokens_per_char badly under-counts CJK, over-counts nothing).