2026-08-11 — The manual nag muted to once per session

Settled after the fleet observation: with several same-build servers reachable from one conversation, or across an app restart, the per-response ;; manual unread line was almost always inaccurate — the manual is per-conversation, the read-mark per-session, and the server structurally cannot see conversation state (sessions don't span reconnects; a mid-conversation instructions swap would bust the client's prompt cache, so nothing else delivers late corrections either). Three findings drove the shape chosen. The repeating nag cannot do its apparent job: it rides eval responses, so the first eval it could inform is the second — the pre-first-eval channels are the instructions and the eval description, which already carry the directive with the conversation-scoped conditional. Its real cost is footer credibility: the footer is where truncation, eviction and settled news are announced, and a line that is usually wrong trains tenants to skim the channel that must be believed. And dropping it entirely was rejected for one asymmetry: instructions delivery is client-dependent, while the footer is in-band and always arrives — so a single first-response line is the one server-controlled channel guaranteed to reach a genuinely fresh tenant. Hence: once per session, worded as fact with the judgment left where the knowledge is (new session — (__r/manual) unless this conversation has), burned only when the response carrying it actually goes out — a client that timed out was never nudged, the same delivery discipline as every mark.