#624 cc-context-monitor measured role-cap version (floor+EOQ headroom) - successor to #617 seed caps

#624 · llmmsg-srv · child of #617

cc-context-monitor measured role-cap version (floor+EOQ headroom) - successor to #617 seed caps

Ref: #624 (#624)
Project: llmmsg-srv
Parent: backlog #617 Fleet context-cost rollout: send-keys /compact watchdog + aro_leave-on-idle
Status: backlog
Priority: low
Type: feature
Assigned: bin-whey-cc coder
Created by: —
Created: 2026-06-03T16:14:34.037Z
Updated: 2026-06-03T16:14:44.321Z

No sub-items.

No questions.

2026-06-03T16:14:34.037Z · created

wi cli; parent=#617
2026-06-03T16:14:44.321Z · note

Spec (bin-whey, nw-whey-approved 2026-06-03): replace v2.13 SEED static caps (pm170/coder160/db160/audit180/researcher150k) with COMPUTED per-agent cap = clamp(floor + sqrt(2*Delta*Ccompact/rate), floor+30k, min(250k, model_window-13k)). floor = rolling MEDIAN of last ~5 post-compact ctx troughs per agent (lowest ctx within ~3 turns of a sharp intra-session ctx DROP), derived from context_samples (no schema change). Delta = per-turn ctx growth = (ctx delta)/(cum_turns delta) over consecutive same-session samples. rate = cache-read $/tok (opus 0.5e-6/sonnet 0.3e-6/haiku 0.08e-6) - cheaper models get MORE headroom. Ccompact = 0.75 flat seed, later measured + model-specific. CONSTRAINT (nw-whey): H* is turn-frequency-independent - measured caps must give high-freq roles (pm/coder) MORE headroom not less; must NOT tighten them. Seeds are interim until this lands.