f32105b3b9
Some checks failed
Contributor Attribution Check / check-attribution (pull_request) Failing after 34s
Docker Build and Publish / build-and-push (pull_request) Has been skipped
Docs Site Checks / docs-site-checks (pull_request) Failing after 2m44s
Nix / nix (ubuntu-latest) (pull_request) Failing after 7s
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Successful in 31s
Tests / e2e (pull_request) Successful in 3m1s
Tests / test (pull_request) Failing after 39m39s
Nix / nix (macos-latest) (pull_request) Has been cancelled
Instead of compressing at a hardcoded 50% of context length, the threshold now adapts to model capacity: - 500K+ context → 75% threshold (large-context models breathe) - 200K-499K → 65% - 128K-199K → 55% - < 128K → 50% (unchanged default, backward compatible) Impact: Claude Opus (1M context) gets 750K working tokens instead of 500K. MiMo v2 Pro same. Small models unaffected. Explicit threshold_percent parameter still works (overrides adaptive). update_model() also recomputes adaptive threshold. Research: See ~/.timmy/research-backlog.md item #4.3 (Ratio: 4.0) Paper refs: KIVI (2402.02750), SnapKV (2404.14469), Self-RAG (2310.11511), Long Context vs RAG survey (2407.16833)