Finding / hermes-agent

2026-10-13-hermes-agent-compaction-on-1m-window-models-moved-from-a-256k-cap-to-half-the-window-in-v0-21-6-and-a

Compaction on 1M-window models moved from a 256K cap to half the window in v0.21.6, and a seeded 256000 cap is migrated away. Since #115986 every context window above about 341K compacted at 256K tokens, so a 1M-window model compacted at a quarter of its window and a compression.threshold above that was ignored. v0.21.6 makes the absolute cap opt-in again and lets the ratio decide, which at the default 0.50 means compaction near 500K tokens on a 1M model. Config v47 deletes threshold_tokens from config.yaml only when it equals exactly 256000, the value the seeder and hermes doctor --fix wrote; any other value survives.

Channel: tagged-release. Half: both. Date: 2026-10-08.

Operator consequence: Long sessions on 1M-window models carry up to about twice as many input tokens per request before compaction on v0.21.6+, on top of an unlimited default turn count. If spend matters, set compression.threshold_tokens to an explicit value other than 256000 (250000 keeps the old behavior) so the migration leaves it alone, or lower compression.threshold, and re-baseline cost comparisons made on earlier tags.

Receipt

Finding metadata

Run: 2026-10-13-profile-refresh-2026-09-21_2026-10-13-frontier-v0

Finding ID: 2026-10-13-hermes-agent-compaction-on-1m-window-models-moved-from-a-256k-cap-to-half-the-window-in-v0-21-6-and-a

Profile citations

Source links

Primary links, including exact changelog lines when available.