updates

Flipped the assistant into a mode where every conversation goes to a human.

A real art direction for covers, a two-track headline contract, and language selection that offers instead of hijacks.

Moved the assistant's judgement calls into lookups, fixed language and formatting leaks, and started recording decisions properly.

Stopped writing for search, fixed how evergreen guides are marked up, and consolidated the review pass.

Bounded retries for failed blobs, a write-scope hole closed in the doctor tools, and driver docs corrected to match the code.

A new course term went live, and the assistant learned to acknowledge promotions it isn't allowed to offer.

Automated the daily publish path end to end, and hardened the review pass around Spanish register and flow arithmetic.

A reliability sweep across startup, sessions, and the pipeline, plus TONU/diversion outcomes and a build-test cycle cut in half.

A new morning cohort (and a correction), a landing page converted to a waitlist, and softer handling when the assistant has to say no.

Quality-control week: review-pass hardening, Spanish translation fixes, and tighter sourcing.

Fleet lists got real pagination and filter routing, expense review got faster, and a batch of CI and MCP fixes shipped.

Reply language moves out of the prompt and into code, stale conversation history gets dropped, and transcripts feed a knowledge pass.