Cinderella: what shipped, and what is next A technical update since the last one. She has a proper help command. Ask in any phrasing, in English or German, or send /help. The capability list is generated from the active intent catalog rather than written by hand, so disabling a plugin removes it from the help text automatically and a new plugin appears without anyone editing anything. help consent and help prices give topic detail. Building that surfaced a resolver collision worth mentioning: help consent was resolving to a price lookup, because help reads as a plausible asset ticker and scored higher than the help intent. Fixed by forcing an instruction that literally begins with help to the help intent, rather than weakening the price resolver and causing a different problem elsewhere. The consent wording now states the terms before you agree, not after. Three properties, all in the prompt: publication is forward only, so nothing you said before opting in is ever included. It stays public until you take it back. And taking it back is final, all at once, with no way to restore it afterwards. Opting in again later starts fresh from that moment. That last property was not previously stated, and it is the one people would not expect. Your questions are archived alongside her answers. Until recently, any message the interaction layer processed as an instruction was never stored at all, so the public archive showed her answers with the questions missing. She appeared to be answering nobody. Question and answer are now treated as a pair: an answer is published only if its question is, and revoking removes both halves. Consent commands themselves stay out, since they are plumbing rather than conversation. Price lookups, on a plugin framework. Ask her what something is worth, or convert between two assets, in natural language. Three data providers behind one interface with automatic failover, attribution bound to whichever provider actually answered, and symbol-to-asset mappings resolved once and then pinned permanently rather than re-resolved on every request. That pinning matters more than it sounds. Provider search rankings shift, so re-resolving each time means the same question can quietly return a different token's price on a later day. One case in particular: Ethereum HEX and PulseChain HEX share an identical contract address, because PulseChain is a state fork. An address-only lookup returns a roughly 2.4x wrong price with no error at all. The plugin framework is the more durable part. A disabled plugin does not merely go quiet, it leaves the intent catalog entirely, so nothing can resolve to a capability that is switched off. Published images are stripped of metadata. Every image served from the public archive comes from a derivative with EXIF, IPTC and XMP removed. Phone photos routinely carry GPS coordinates and device serial numbers, and consenting to publish your words is not consenting to publish your home coordinates. The gate fails closed: a file that cannot be stripped is withheld rather than published unstripped. What is next Choosing what revocation does. Rather than a single irreversible action, you will pick: hide, which takes your words out of the public archive but keeps them so you can restore them later, or delete, which removes them for good. Deleting will require typing the word itself, so a reflexive yes cannot erase an archive. Alongside that, evidence holds. Content reported as illegal is held from destruction until it has been reviewed. The important detail is the separation: a hold prevents destruction, never hiding. You can always take your content out of public view immediately, with no review and no delay. Only the physical deletion waits. Reports are anonymous, so without that separation anyone could make someone else's content permanently undeletable simply by reporting it. Video links will play in the stream, with a design constraint worth stating: nothing loads from any third party until you click. The thumbnail is fetched once server-side and served from our own domain, so opening the archive produces no requests to external services at all. Alongside it, a consent banner with separate categories, where rejecting is exactly as easy as accepting. And a private channel. Right now everything she says is said in front of everyone, which is wrong for anything personal. Asking about your own archived content should not announce it to the room, and choosing whether to hide or delete should not be negotiated in public. That work includes a way to prove a private conversation belongs to a specific group member, because she must never change anyone's consent on the strength of an unverified message.