CortexDB 0.12.7 — hosted plans: answers have their own allowance, memories count by size

0.12.7 is 0.12.6 with two changes to hosted plans (CORTEX_PLANS=1). An on-prem server, where plans are off, behaves exactly as 0.12.6: upgrading is a drop-in replacement.

What changed (hosted plans only)

• Answers have their own monthly allowance. /v1/answer and /v1/compose (the requests that reach the answer model) now count against answers_month, not against recalls: free 100, starter 2,500, pro 15,000, enterprise unlimited. Past it they answer 402 PLAN_LIMIT_REACHED with details.limit: "answers_month", used, max, plan, subtree, upgrade_url and resets_at. Recalls (/v1/recall, /v1/recall/stream, and an answer with skip_answer_llm, which never reaches the model) keep their own limit. A request counts once however many passes it makes, and one refused before its model (no llm.invoke, no answer renderer) counts nothing.
• Memories count by size. A memory is up to 2,000 characters of its text (its text, a blob's transcript, otherwise its serialized content); a longer one counts once per started block, so 5,000 characters count 3. Writes, bulks (all or none) and the storage recount use the same rule, so an account's count is the same after a restart, and a forget gives the units back. Existing data is re-counted the same way.
• The plan report (GET /v1/plan, GET/PUT /v1/admin/plans/{subtree}) adds limits.max_answers_month, usage.answers_this_month and memory_unit_chars (2000). An operator can override max_answers_month like the other limits.

Upgrading a hosted server from 0.12.6

Nothing to configure. The month's answer count starts at zero; answers already made this month were counted as recalls under 0.12.6 and stay there. The next recount of each account (within 10 minutes, or at its first write) measures its memories by size, which can raise the count of an account that stored long memories.