Limitations

Headroom's lossless fold runs first

Headroom applies its byte-exact lossless fold (and, for logs and search results, relevance_split) before the external compressor hook, and returns early when either succeeds. Blocks that fold, such as logs with repeated lines, never reach Winnow. This is pinned by tests/test_router.py::test_known_gap_lossless_fold_preempts_external, which will start failing if Headroom changes the order.

Passthroughs in Headroom 0.39.1

When an external compressor declines a block (compressed=False), Headroom 0.39.1 still adopts the unchanged block as the result and skips its built-in compressors for it. A one-line fix is proposed upstream (upstream/router-respect-passthrough.patch); until it lands, test_router_falls_back_when_backend_unavailable is a strict xfail that will flip automatically.

CPU performance

On CPU the pooled model's budget is 2,048 tokens and the highlighter's is 512. Larger outputs go to Headroom's own compressors. A GPU is recommended.

Scope

Winnow handles logs, search results, diffs and plain text. JSON, source code, HTML, CSV and config files are left to Headroom's structure-aware compressors, which already handle them well.