Repository navigation
0.16.0
·
31 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
Added
- The optional assistant's core, behind a flag (#17, ADR 0009). Nothing changes unless an operator sets
DINNERBELL_ANTHROPIC_API_KEYand addsassistanttoDINNERBELL_FEATURES; then Admin gains an Assistant tab. There the admin turns the assistant on, sets the household's Monthly cap ($1 to $100 a calendar month in the household's time zone, $5 by default) and Conversation hours (0 to 168, 24 by default), and sees this month's usage in total and per member. Each change is audited ashousehold.assistant_updated. Members talk to it in Slack and on the web's Assistant page (below). This release adds what they build on:- Before each request to Anthropic, Dinnerbell reserves the most the request can cost and refuses it if that would pass the cap, then settles the reservation to what the request cost, cache tokens included. At the cap, the assistant says it's unavailable until next month.
- Conversations are stored with members as "Member 1", "Member 2", and so on, and deleted once Conversation hours pass, every 5 minutes, or at once when the admin lowers the setting. No names, emails, Slack IDs, preferences, vetoes, ratings, or notes are sent.
- The assistant's one tool so far looks up catalog items by name or tag, acting as the member with a member's permissions, even for an admin.
DINNERBELL_ASSISTANT_MODELpicksclaude-sonnet-5-5(the default),claude-opus-5-5, orclaude-haiku-5-5; Dinnerbell refuses to start on a model it has no prices for.dinnerbell scrub-testturns the assistant off and deletes every conversation.
- The assistant's candidates, drafts, and Confirm, still behind the flag (#17). Four more tools, each acting as the member with a member's permissions, even for an admin:
- Find candidates, for a call that's voting: what passes ranking's filters for everyone who's in, in name order, optionally narrowed by days since eaten, cuisine, or mode. Reasons come only from variety, novelty, suggestions, the usual mode, volunteer fit, and meal fit, with members as "Member 1" and so on; never hints, favorites, or what the filters removed. Before voting, the assistant says to ask once voting opens.
- Draft a proposal of a candidate, draft an offer of a catalog item, and draft catalog items in the catalog file format. A catalog draft drops entries the catalog has already and lists them, and strips
cooksandactive. - Each draft says what Confirm does, such as join late, take the role the mode needs, or replace a scheduled call, lasts 30 minutes, and only its member can confirm it, once. Confirm runs the same service as the form, which checks everything again, so a stale draft gets the usual error and nothing changes. A catalog draft's Confirm opens the import preview, which only adds, even if a matching item appeared meanwhile.
- The tools pick the call the member names, then the conversation's, then the only open one, and ask when several are open.
- Expired drafts are deleted every 5 minutes, and
dinnerbell scrub-testanddinnerbell reset-historydelete them all. The web page and Slack have the Confirm and Cancel buttons (below).
- The assistant on the web: the Assistant page, behind the same flag (#17). With the key, the flag, and Admin's switch on, each active member and admin gets an Assistant link and a private conversation of their own; guests don't. A reply appears under the message without reloading the page, and the form works without JavaScript too. Replies show bold, italics, lists, and code, and their links aren't clickable. Each draft is a card that says what Confirm does, with Confirm and Cancel; a proposal or offer shows the service's result or its refusal on the card, and a catalog draft opens the import preview, which only adds. A draft that expired, was used, or isn't yours says so. Start over deletes the conversation and cancels its waiting drafts. At Conversation hours 0, the page says nothing is kept and shows only the latest exchange. Otherwise the link is hidden and the Assistant pages answer 404, as Admin > Assistant does.
- Confirm the assistant's drafts in Slack (#17). Each draft comes as a card after the reply, with what Confirm does, and Confirm and Cancel. In a channel thread the card shows only to the member who asked; in a direct message it's an ordinary message in the thread. Confirm runs the same service as the form, as that member, and the card then says what happened, or the service's reason it can't, such as a call that's no longer voting. A draft that's expired, used, or someone else's says so and changes nothing. A catalog draft's Confirm gives a Review the import link to the web's add-only import preview, signed for that member and good for 24 hours, as long as Dinnerbell keeps a confirmed catalog draft; nothing is saved until they select Apply there. While the assistant is off, the buttons say "The assistant is turned off." and change nothing. A word that only starts like
@here,@channel, or@everyone, such as "@hereford", now reads as written in replies, and the note that Dinnerbell sees only mentions comes once to each member whose replies there were private, and once to the thread with its first public reply. - The assistant drafts catalog items from a menu link, behind the same flag (#17). Give it a link to a restaurant's menu or a recipe, and it drafts catalog items for Confirm, which opens the import preview, as other catalog drafts do. Dinnerbell fetches the page under the same rules as every menu link (HTTPS, public addresses only, 2 MB, 10 seconds), and sends only its text, up to 60,000 bytes, to a separate drafting call with no tools, reserved and settled under the monthly cap like any other call. A message reads at most 2 links. The page is treated as data, never instructions: it sits between randomly named tags, the answer is only read as a catalog file, and the draft keeps at most 20 entries and drops every link and note the page gave. A draft of one place gets the member's own link, or the page's address when the link redirected within its own site. The assistant sees just how many items the draft adds and which ones the catalog has already.
- Talk to the assistant in Slack, behind the same flag (#17). With the key, the flag, and Admin's switch on, a member can message Dinnerbell directly, or mention @dinnerbell in the household channel, and Dinnerbell answers in that message's thread. A direct message thread is private to its member. A channel thread is shared, each message acts as its author, and Dinnerbell's first reply there says it sees only messages that mention it. Dinnerbell answers active members only, and ignores bots, edits, and mentions outside the household channel: nothing from anyone else is stored or sent. Replies can't mention anyone or ping the channel, and hold no links: any address in one is written so Slack doesn't link it (
hxxps[:]//example[.]com). Each member has one message answered at a time; another meanwhile gets a note that Dinnerbell is still on the last one. A direct message with a file gets "I can read only text messages." App Home shows a line saying members can message or mention Dinnerbell while the assistant answers them. Slack needs the new assistant variant of the app manifest,deploy/slack-app-manifest.assistant.yaml(deploy/slack-app-manifest.assistant.test.yamlfor a test instance); the default manifests don't change. - The assistant answers within 100 seconds, and keeps who a Slack message mentions (#17). A message that takes longer gets "I ran out of time on that one. Try asking for less at once." instead of running on after the web page stops waiting, and nothing of it is kept. In Slack, a mention of an active member reaches the assistant as that member's "Member 2", so "can <@kim> cook?" keeps its meaning, and the reply names Kim; anyone else still goes as "@Someone".
- The assistant answers questions about meal history, behind the same flag (#264). Ask "What's our most-ordered pizza place?" or "When did we last have Thai?" in any conversation, and the assistant looks it up in the meal log: each place or home meal the household ate, how often, and when last, narrowed by name, tag, kind, or dates. Each comes with its household hint, such as "Well liked", the same words every member sees on Catalog, under the same thresholds. It never says who ate, cooked, ordered, or rated, and sends nobody's own ratings or notes (ADR 0009, amended).
- The assistant shows that it's working in Slack (#265). As soon as Dinnerbell takes a message to answer, the thread shows "Dinnerbell is thinking…" with a few rotating lines. Where Slack won't show that status, Dinnerbell reacts to the member's message with ⏳ instead. Either one goes when the reply posts, and after an error or the out-of-time line. The indicator never delays or blocks the reply, and messages Dinnerbell ignores show none. The assistant manifests add the
reactions:writescope for the reaction. - The assistant drafts call actions, in any conversation, behind the same flag (#251). In a DM, on the web, or in a channel thread, a member can ask the assistant to answer in, out, or maybe; join late; vote or take a vote back; suggest a catalog item; take or give up a role; take back an offer they made; invite a guest, saved or new with the dietary needs they name, and answer for a guest they invited; or start a call, or start a called-off one again. Each change is its own draft, so "I'm in, put me down for pickup, and vote for the noodle place" gives three cards, and each Confirm runs the same service as the button on the call, as the member with a member's permissions, even for an admin. A draft says what Confirm does, such as joining late first or withdrawing role offers on an Out, and the assistant refuses what the call wouldn't allow, such as answering after answers closed, taking back someone else's offer, or answering for someone else's guest. Guests are named only as the member wrote them. A new Look up a call tool gives the assistant the call's basics, the member's own answer, and the picks with whether the member votes for each; never anyone else's answer or vote. In a channel thread, the cards and their results show only to the member who asked.
- The assistant drafts changes to a member's own preferences, feedback, and away days, in private only, behind the same flag (#253). On the Assistant page or in a direct message with Dinnerbell, never in a channel thread, a member can ask "Veto mushrooms", "Rate last night's dinner: hated it, it was cold", or "I'm away Thursday through Saturday", and get a draft for each. Confirm runs what Preferences and the feedback form run, as the member with a member's permissions, even for an admin, and never Admin's path to someone else's preferences. A draft says what Confirm does, such as what feedback it replaces or that a veto keeps something out of every call they're in, and the assistant refuses what the form would, in the form's words, such as feedback after its 7 days or on a meal they didn't eat. If the member changes the same thing elsewhere first, Confirm says so and changes nothing. Two new lookups give the assistant the member's own levels and away days, and the meals they ate that still take feedback, with their own answers. Only these tools send a member's own preferences, ratings, feedback, or away days to Anthropic, only when that member asks, and never anyone else's: not who else ate a meal, who cooked it, or how anyone else rated it. Admin > Assistant says so. A cap too low for even one request, such as $1 with
claude-opus-5-5, where one private message's worst case is now just over $1, gets a truthful reply that the cap is too low for the model, not that the month's limit is reached, and Admin > Assistant notes it with the least cap that covers the smallest message, saying long conversations need more. - The assistant's longer Slack replies appear as they're written (#266). In a direct message, a reply of more than a couple of lines streams into its thread line by line, through the same defanging as any reply: no link, mention, or ping can get through however the text is cut. A streamed reply is standard Markdown, so bold, lists, and code show, and a
<shows as‹. Short replies, every reply in a channel thread, and the web page don't change, and neither does what's stored. If Slack won't stream, the reply is posted as before, without repeating what the member already saw. The assistant is also asked to write no text before it calls a tool, so a tool round's preface doesn't show. No new scope or manifest change. - The assistant coordinates a call, in private only, behind the same flag (#252). In a direct message with Dinnerbell or on the Assistant page, never in a channel thread, the coordinator can ask the assistant to lock a pick, break a tie, or shuffle the picks while voting; once it's decided, change the plan as Adjust does (the meal time, a catalog place, the way, and who cooks, orders, or picks up); hand off coordinating; mark the meal done once it's meal time; or call the call off, with a reason. "Lock the taco place, @sam picks it up" gives one card that locks it and assigns the pickup together, so Sam hears once, and if the roles can't be assigned, nothing is locked. The member who started an ad-hoc call can call it off too, until it's decided. Confirm runs the same use case as the call page or Slack's button, as the member with a member's permissions, so the coordinator's own checks decide: an admin who isn't the coordinator is refused, as any member is, and uses the web instead. A change of plan that someone made elsewhere first is refused rather than written over. In a channel thread, the assistant says to ask in a direct message. People go to Anthropic only as "Member 2" and so on: a new lookup gives the coordinator's assistant who's in, with the roles they offered, how many guests are in, which picks are tied, and the plan; never anyone's vote, maybe, or out. The coordinator's cards and Confirm's results, which only they see, show names; the model reads only labels, and a refusal that would name someone, such as a cook list's, comes to it with the label. The seven new tools make each private message's worst-case reservation about 15% larger, about $0.31 with
claude-sonnet-5-5. - The assistant edits the catalog and the meal log, in any conversation, behind the same flag (#254). On the web, in a direct message, or in a channel thread, a member can ask the assistant to edit a catalog item, such as "Mark Siam Garden as delivery only", take one off the catalog or put it back, log a meal they had without a call, such as "We had Taqueria for lunch yesterday", or edit or delete a meal they logged. Each is a draft, and Confirm runs what the Catalog item form or History runs, as the member with a member's permissions, even for an admin: an admin's own History powers, such as changing someone else's meal, stay on the web, and renaming or merging tags is never a tool. A catalog edit's card is a before-and-after diff, one line for each field Confirm changes, with the new value first and in full, in Slack too, and nothing else changes. Cooks, and who ate, are named only by the conversation's labels, such as the member's own or a Slack mention's, and resolved to active members of the household when the draft is made, so the model can't make up a cook. If the item or the meal changed elsewhere after the draft, Confirm says so and changes nothing. A new Look up my meals tool gives the assistant the meals the member logged, with who ate as labels in private only: in a channel thread, whose reply posts for everyone, it leaves out who ate. An item's link, order link, and notes as they are go to nobody but the member's card. In a channel thread, the cards and their results show only to the member who asked. The three new tools, and the prompt's lines about them, make each message's worst-case reservation about 6% larger in private and 7% in a channel thread: about $0.33 for a private message with
claude-sonnet-5-5(from $0.31), and $1.23 withclaude-opus-5-5(from $1.16). A channel thread's message withclaude-opus-5-5now reserves about $1.03, so a $1 cap with that model answers no message at all, and Admin > Assistant now says so.
Changed
- Enter sends a message on the Assistant page (#285). On a keyboard, Enter sends what's in the message box, as Send does, and Shift+Enter starts a new line. Enter that picks an input method's candidate doesn't send. On a touchscreen, Enter still starts a new line, and Send sends.
- Shuffle proposals is easier to find on the web (#267). On Today and the call page, the button now sits under the proposal list, where Slack puts it, with its "N left" note, and has a border so it reads as a button, while Vote still stands out. Who sees it is unchanged: the coordinator and admins, while a shuffle is left.
Fixed
- The assistant's formatting shows (#284). Replies came with their Markdown markers showing, such as
**Tag**and-lists, in Slack and on the web. Slack now renders a posted reply's Markdown, as it already did for a DM reply as it streamed. The Assistant page renders a small, safe subset: bold, italics, lists, and code. Links still aren't clickable anywhere, and the reply can't mention or ping anyone. The assistant is asked to use only bold, italics, and lists. - The assistant answered nothing (#280, #282). Every request failed with a 400: first because Anthropic takes at most 20 strict tools in one request and the web page and direct messages offered 31, then because the grammar Anthropic compiles from even 19 strict tools was too large. No tool is strict now. Every tool already checked its own input, and a test breaks each field of every tool in turn to prove it. A failed request's log line now has Anthropic's error message in
detail, cut at the first quote mark so no quoted value from the request is logged. - CI only: when the runner can't install podman-compose with apt, the
composejob installs the same version, 1.0.6, from PyPI, without Ubuntu's patches but with the python-dotenv and PyYAML versions Ubuntu 24.04 ships (#281). A new step fails the job if podman-compose 1.0.6 isn't on the PATH, sincescripts/check-compose.shskips it when it's missing. On 2026-11-11 the runners' DNS failed for everyubuntu.comhost for hours, andpublish, which waits forcompose, couldn't run. Each apt run is capped at 2 minutes, and the job's limit is 15 minutes.
Upgrade notes
- Migration 0031 adds two empty tables,
assistant_usageandassistant_conversation. It changes no existing row, and nothing reads the tables unless the assistant is configured. A downgrade to 0.15.0 drops them, with any usage and conversations in them. - Migration 0032 adds one empty table,
assistant_draft, for the assistant's drafts. It changes no existing row, and nothing writes the table unless the assistant is on. A downgrade to 0031 drops it, with any drafts still waiting for Confirm. - Migration 0033 adds
assistant_draft.confirmed_at, empty for every draft saved before it. A catalog draft confirmed in Slack before the upgrade loses its Review the import link; ask the assistant again. A downgrade to 0032 drops the column. - Migration 0034 rebuilds
assistant_draftso itskindaccepts the call actions (#251), keeping every draft and its IDs. A downgrade to 0033 deletes call-action drafts still waiting for Confirm. - Migration 0035 rebuilds
assistant_draftagain so itskindaccepts drafts of a member's own data (#253), keeping every draft and its IDs. A downgrade to 0034 deletes those drafts still waiting for Confirm. - Migration 0036 rebuilds
assistant_draftagain so itskindaccepts coordinating drafts (#252):decide,shuffle,plan,hand_off,mark_done, andcall_off. It keeps every draft, its IDs, and the AUTOINCREMENT high-water mark, so an old Confirm button never reaches a newer draft. A downgrade to 0035 deletes those drafts still waiting for Confirm. - Migration 0037 rebuilds
assistant_draftagain so itskindaccepts catalog edits and the meal log (#254):item_edit,item_active,meal_log,meal_edit, andmeal_delete. It keeps every draft, its IDs, and the AUTOINCREMENT high-water mark, so an old Confirm button never reaches a newer draft. A downgrade to 0036 deletes those drafts still waiting for Confirm. - New dependency: the image now includes Anthropic's
anthropicSDK and its dependencies (Apache-2.0, MIT, and BSD). Dinnerbell sends nothing to Anthropic unless you set the key and the flag. - The cap is per database: production and a test instance each keep their own. If you turn the assistant on, also set a monthly spend limit on the key's workspace in the Anthropic Console as a backstop, and give the test instance its own key.
- Slack manifest, only if you turn the assistant on: before you turn on Admin's switch, paste
deploy/slack-app-manifest.assistant.yamlunder App Manifest, set your hostname again, and reinstall the app (on test,deploy/slack-app-manifest.assistant.test.yaml). It adds a writable Messages tab, theim:history,app_mentions:read, andreactions:writescopes, and themessage.imandapp_mentionevents. If you applied an assistant manifest from an earlier build, such as on a test instance, paste it again and reinstall, so the app getsreactions:write; until then, a message gets only the status, where Slack shows one. Without the assistant, keep the default manifest. See "The assistant" in docs/configuration.md. - Keeping the key in a file:
compose.secrets.yamldoesn't mountDINNERBELL_ANTHROPIC_API_KEY, so installs that keep secrets in files need no new file. To keep the key in one, see "The assistant" in docs/configuration.md.
Image
ghcr.io/dsmorgan/dinnerbell:0.16.0@sha256:36a1404e8e468c2360067c70a96e97d25488b03b5d83f03d80551765a0820d4e