CHANGELOG.md
Rama main · réplica de lectura del repositorio original.
# Changelog
All notable changes to killBottleneck. Dates are the release date of the tag.
The version you are running is shown in the **About** dialog; the same string is in
`KB_VERSION`. Upgrading is always `git pull && docker compose up -d --build`
(see [Updating](https://killbottleneck.com/guide/updating)) — read the **Upgrade notes**
below before you jump several versions.
---
## v1.0.1 — 2026-11-08
**The assistant on your own OpenAI key: reasoning models without extra settings, reminders and recurring events**
**Upgrade notes:** no migration. If you added `KB_CHAT_OPENAI_EXTRA={"reasoning_effort":"none"}` to make the
assistant work on OpenAI, you can keep it (it still takes precedence) or delete it.
- **Reasoning models work with the assistant out of the box** (`openai` provider, e.g. `gpt-5.6-luna`). OpenAI
refuses a reasoning setting other than `none` together with tool calling, so every assistant turn ended with
*error 400*. The app now retries with `reasoning_effort: none` and then without the field, and remembers per model
what worked, so later turns go through on the first call.
- **Drawing through OpenAI defaults to `gpt-image-1-mini`** instead of `gpt-image-1`, which many new API keys no longer
have (every picture failed at once). Set `KB_IMAGE_MODEL` to choose another model. The error message on a refused
request no longer shows doubled braces.
- **Reminders without a project:** "Remind me tomorrow at 8 to call the bank" now becomes a calendar event with a
reminder, not an idea in the inbox.
- **Recurring events:** "Every Monday at 9 we have a team meeting" becomes a series of events in the calendar (up to
30); when you don't say until when, the assistant usually asks first. A repeating *step* in a map keeps its own rhythm as
before.
- **New guide: [Your own OpenAI key](https://killbottleneck.com/tutorials/openai)** — the assistant and drawing on a
self-hosted server with one OpenAI key, which models to pick, what we measured and what to do when something fails.
---
## v1.0.0 — 2026-11-07
**The first stable release — and killBottleneck is open source under the MIT license**
**Upgrade notes:** one data migration fills in the language of existing accounts (`users.language = cs` wherever it
was empty) so that nobody's e-mails or notifications change language. This is the first release the stable update
channel has offered since v0.33-beta — if you jump from an older beta, read the *Upgrade notes* of the versions in
between as usual (`git pull && docker compose up -d --build`, a backup first).
- **License: MIT.** The whole product — server, web client, MCP server, desktop app — is released under the
[MIT License](./LICENSE). Download it, run it, modify it, host it for others, build on it. The beta releases (v0.x)
stay under the Sustainable Use License; the change applies from v1.0.0 on.
- **The name and the logo stay with the authors** — see [TRADEMARKS.md](./TRADEMARKS.md). A fork needs its own name;
"based on killBottleneck" is welcome. The commercial-license offer (`licence@`) is gone from every text.
- **End of the beta.** The website, the README and the app no longer say "public beta"; the `KB_UPDATE_PRERELEASE`
switch now means release candidates (`-rc`).
- **English first.** `docker-compose.yml`, `.env.example` and the `Dockerfile` are commented in English only; new code,
comments and logs are written in English; `AGENTS.md` describes the repository for people and AI agents. A regression
test (`en-first`) keeps the amount of Czech in comments and logs from growing.
- **Behaviour change — fallback language is English.** A user whose account has no language set now gets English
e-mails, notifications and assistant texts instead of Czech, and so do answers to visitors who are not signed in
(closed registration, public map links) unless their browser asks for Czech or Slovak. Existing accounts are not affected: the migration above
writes `cs` into every empty `language` field, which is exactly how the server treated them before. An account created
without a language (API, MCP) gets English server texts until its owner signs in through the web app, which then
stores the browser's language (Czech/Slovak → Czech, everything else → English) as before.
- The *About* dialog describes the MIT license and links to its full text; `NOTICE.md` and the third-party list are
updated accordingly.
**Fixes from the feature analysis of 5 Nov 2026** (behaviour defects verified live, no new decisions):
- **"Delete selection" no longer loses data.** Bulk deletion skips goals whose task was assigned by someone else
(only the assigner or the project owner may remove them) and says so, exactly like the trash button and the
Delete key. Before, the server refused the whole save, the app stayed silent and later work was lost on reload.
- **Lite quick add offers only projects you can write to** (owner, named "edit" share, team edit). A read-only
project ended with "The requested resource wasn't found".
- **Archiving by a non-owner is refused out loud** (400 with a message) instead of a silent 200 with no effect.
- **Transferring project ownership (API for now, no button yet):** `POST /api/kb/maps/{id}/transfer` (owner or administrator; the previous
owner stays a co-admin with edit rights) and, before deleting an account, *User administration* offers to take
over the user's projects (`GET /api/kb/users/{id}/maps`, `POST /api/kb/users/{id}/transfer-maps`) — deleting an
account used to delete its projects by cascade without warning. A bare `PATCH` of `owner` now derives
`owner_email` and the share lists the same way.
- **MCP: new `update_map` tool** (title, description) — renaming a project through an agent used to change only the
apex text while the map list and the database kept the old title. 35 tools now; the HTTP `/mcp` catalog and the
stdio server stay 1:1.
- **Welcome map of a member:** the root goal reads "Welcome to killBottleneck" like the title (it said "Get going on
your own").
- Small: the timer is shown on the apex for readers too; the Google button is hidden on the sign-up page in the native
shell (as on sign-in).
---
## v0.101-beta — 2026-11-06
**License housekeeping from the third code review — nothing copyleft, every notice in order**
**Upgrade notes:** one data migration (`1793956000_canvas_texty_vlastni`) rewrites the block texts of the two system
templates *Business Model Canvas* and *Lean Canvas* — only if they still carry the original seed wording; a template an
admin edited is left alone (and logged), personal copies are untouched. Rebuild the image as usual.
- **PDF fonts renamed.** The Latin subsets of Liberation Sans / Liberation Serif used for invoices, Office PDFs and
PDF text fixes are now called **KB Sans / KB Serif** (`frontend/public/fonts/pdf/KBSans-*.ttf`). The OFL reserves the
name "Liberation" for unmodified fonts; the files now ship the upstream OFL text instead of a Debian copyright file
that mentioned GPL-2+. Generated PDFs look exactly the same.
- **Business Model Canvas and Lean Canvas** credit their authors (Osterwalder / Maurya, CC BY-SA 3.0) in the template
description; the Lean Canvas block descriptions are reworded in our own words instead of the original's.
- **License texts travel with every copy:** the Docker image serves `/LICENSE.txt` and `/THIRD-PARTY-LICENSES.txt`,
and the *About* dialog links to both (also from the Android app, where the files live on the server). The About text no longer promises "no strings" — it says what the Sustainable
Use License allows (own use, commercial too; no hosted resale, no white-label).
- **`THIRD-PARTY-LICENSES.md`** regenerated: dual-licensed packages are marked with the option we use and the GPL text
of jszip is no longer reproduced; Plus Jakarta Sans, the KB fonts and shadcn/ui are listed; the generator's allowlist
no longer lets "(GPL-3.0 AND MIT)" through (`scripts/licence-pravidla.mjs`, shared with the test). `NOTICE.md` is a
short, current notice instead of a stale package table.
- **Website:** the Geist font ships its OFL text and is listed in `docs/THIRD-PARTY.md`.
- **MCP package** `killbottleneck-mcp` 0.101.0 — the README lists `generate_image`, the version follows the app again.
- Tests: `licence-soubory` (28 checks over the tree), `licence-instance` (instance serves the texts, templates seeded
with the new wording), `upgrade-canvas-texty` (existing instance gets the new texts, personal copies stay).
---
## v0.100.1-beta — 2026-11-05
**A weekly picture count instead of credits**
**Upgrade notes:** no migration. Nothing changes unless you set the new variable.
- **`KB_OBRAZKY_TYDEN`** caps the pictures of the whole instance per week (from Monday, like the credit quota; empty / 0 =
no cap). Every successful drawing counts — chat, New image, redraw, Office and Pages; the daily-summary banner does not.
Over the cap the card ends with "All N pictures of this instance for this week are used up…", before any model or
picture service is called.
- **`KB_OBRAZEK_KREDITY=0`** now means pictures are outside the credit quota: neither the drawing nor the prompt that
the assistant writes for New image costs credits, and an exhausted weekly credit quota no longer blocks them.
---
## v0.100-beta — 2026-11-05
**Pictures drawn by the assistant**
**Upgrade notes:** four migrations (collection `ai_images` and its flags, `users.sumar_email`, `daily_summaries.odeslano`) run on
start, no manual step. Drawing is **off until you connect a picture service** — without `KB_IMAGE_*` nothing new
shows up and the assistant's prompts are unchanged. To turn it on with your own service, set
`KB_IMAGE_PROVIDER=openai`, `KB_IMAGE_TOKEN` (API key) and optionally `KB_IMAGE_URL` / `KB_IMAGE_MODEL`
(default `gpt-image-1`, the service must return JPEG); see `.env.example` and the Pictures guide.
- **The assistant draws.** "Draw an illustration…", "make an e-mail banner…": the assistant writes a detailed
prompt for the picture model and a card with the picture appears under the reply (download, redraw). Pictures with
readable text need `KB_IMAGE_TEXT=1` and a second model; short captions work, an occasional typo is fixed by redrawing.
- **Office → Pictures tab:** New picture without a chat (four shapes, seven styles, Painting by default), search, a
viewer with arrows, full screen, download, redraw, pin forever and delete. Pictures used in no document or page are
deleted after 90 days (the viewer says when); used and pinned ones are kept, and a used picture cannot be deleted.
- **Office documents and Pages get pictures on their own:** a title picture and up to three slide pictures in
presentations, a picture on cards, invitations and posters, and up to three illustrations in assistant Pages
(public `/p/` pages included). Without a picture service everything works as before, just without pictures.
- **Daily summary by e-mail** (each user turns it on in Notifications) with the day's banner when pictures are
on; `KB_SUMAR_EMAIL_MAX` caps summary e-mails per instance and day (default 100).
- **API and MCP:** `POST /v1/images` and the MCP tool `generate_image` (34 tools now).
- **Limits:** a picture counts `KB_AI_GEN_VAHA` (default 2) towards the hourly AI limit; `KB_OBRAZEK_KREDITY` sets a
fixed credit price per picture (empty = the model's reference price). Drawing interrupted by a restart ends as an
error with "Try again" instead of spinning forever.
---
## v0.99-beta — 2026-11-05
**Fixes from the third code review (wave A)**
**Upgrade notes:** no migration. The Docker image now includes `sqlite3` (+2.8 MB) because `backup.sh` uses it
for consistent snapshots — rebuild with `git pull && docker compose up -d --build` before your next backup; an older
image makes `backup.sh` stop with a clear message instead of writing a possibly broken archive. Behaviour change:
events in an **archived calendar can no longer be edited or deleted** (unarchive first), the same rule that already
applied to creating them.
- **Backups are consistent while the instance runs.** `backup.sh` takes an `sqlite3 .backup` snapshot of the
databases instead of packing the live WAL files (under write load 4 of 35 such archives were silently broken);
`backup.sh restore` checks `PRAGMA integrity_check` before stopping the instance and refuses a damaged archive.
- **The assistant survives reading large maps.** Tool results stored in the conversation are trimmed to what the
model actually receives; previously the fourth read of a 250-node map hit the 200 kB message limit, the turn failed
with 502 and the user's message (voice note, photo) was lost. A failed save is now logged and reported honestly.
- **Calendars "for selected people only" can be assigned again.** The member list in Organization settings →
Calendars was always empty since v0.83 (`getMembers` returned an object, the list expected an array).
- **"Today / This week" in My Day matches the task list.** The client filter mirrors the server's definition
(plan and deadline, "this week" = 2–7 days), so the chip count and the filtered list agree; the manual filter on the
Tasks page is therefore labelled "In 2–7 days" now (today and tomorrow have their own chips); overdue tasks planned
for later are no longer shown as overdue; the idea buffer is filtered by "Today" too.
- **Timeline:** week 53 of 2026 is one block (ISO year in the week key) and a node's own plan moves the start of its bar.
- **"Today" without a client date uses the instance time zone** (My Day, portfolio, assistant, summaries) instead of UTC.
- **Archived calendars are read-only**: editing or deleting their events returns 400 like creating them did.
- **The "new version" bell no longer repeats** every 30 days on instances that have not updated.
- **MCP via a custom domain:** discovery and `WWW-Authenticate` answer with the address the client used
(slug or custom domain), so connectors created before switching the domain keep working.
- **Invited users get the inviter's language** (or the language sent with the invitation) — invitation and welcome
e-mails are no longer always Czech.
- **MCP package:** the stdio server knows the `offer_assistant` rule action and the `mode` field; its README lists all 33 tools.
- Dev: `cloud/kontrola-env.sh` is green again and runs early in `run-all.sh` (right after the secrets check).
## v0.98-beta — 2026-11-03
**A different skin every day and clearer Organization settings**
**Upgrade notes:** migration `1793260418_user_skin_rotace` adds the field `users.skin_rotace` (on/off, off by default)
and runs automatically on start; no data changes and nothing to configure. Old links keep working: `/admin/users#ai`
opens the *AI* tab and `/admin/users?fakturace=1` the *Subscription and invoices* tab. The Android app shows the new
skin option after its next update. Self-hosters: `git pull && docker compose up -d --build`.
- **A different skin every day.** New switch in *Appearance*: the built-in skins rotate on their own, the same one on all your devices in the same time zone (also in the simplified view and via the assistant).
- **Clearer Organization settings.** The page is split into six tabs (People · Organization and look · Structure and deputies · Shared calendars · Subscription and invoices — just *Invoices* on a self-hosted server · AI) with a tab list on the left; the people table has 7 columns instead of 9 with password reset and account deletion in a ⋯ menu, and phones get a card per person. Nothing was removed and no permissions changed.
## v0.97-beta — 2026-11-03
**Security update — please upgrade**
**Upgrade notes:** the bundled PocketBase moves from 0.39.6 to **0.39.11** (fixes a publicly known bug where a crafted
request could crash the server — GHSA-84vh-m24q-wjjx) and the base image to Alpine 3.22; the download is verified by
checksum during the build. Migration `1791700000_uploads_protected` runs automatically on start (once). **Sign-in rate
limiting is now on** (20 attempts / 60 s, 20 record creations / 5 s, 300 API requests / 10 s, per client IP). If you run
behind a reverse proxy (Caddy, nginx — see [HTTPS](https://killbottleneck.com/guide/https)), add
`KB_TRUSTED_PROXY_HEADER=X-Forwarded-For` to `.env`, otherwise the whole company counts as one client; `KB_RATE_LIMITS=0`
turns the limits off. No data changes. Self-hosters: `git pull && docker compose up -d --build`.
- **Security:** PocketBase 0.39.11 — a crafted request can no longer crash the instance.
- **Security:** sign-in and the API have server-side rate limits per client IP — password guessing is slowed down
(new variables `KB_RATE_LIMITS`, `KB_TRUSTED_PROXY_HEADER`, see the upgrade notes).
- **Security:** the "own AI server" address check could be tricked by a specially formed address into reaching the hosting
provider's internal network; the check now reads the address the same way the connection does.
- **Security:** comments — after posting, the author and the map a comment belongs to can no longer be changed through
the API (only the text can).
- **Security:** API keys — and the OAuth clients of connected AI agents — are revoked when the owner changes their
password, resets it through "Forgotten password", or an administrator resets it. Connect the agent again afterwards.
- **Security:** the legacy `uploads` collection (not used by the app — attachments live in `node_files`) is closed:
its files are no longer readable without signing in and nothing new can be uploaded into it.
---
## v0.96-beta — 2026-11-02
**Office in the assistant: a presentation (PowerPoint or PDF), a greeting card, invitation, poster or certificate (PDF), and every document as PDF or Word**
- **Presentation** — "Make a presentation of project #12 for the client": the assistant takes the facts from the map and writes
the slide outline into **Documents** (editable text: `# title`, `## slide`, `- bullet`, `> speaker note`); the card and the
Documents panel offer **PowerPoint (.pptx)**, **PDF** (16:9) and **Word**. The deck uses your skin colour and the organization logo;
speaker notes go into the PPTX. Ask for changes ("shorten slide 3") — the assistant rewrites the outline, the previous version stays.
(A web page shared by link is still the assistant's Pages; this is a file to download.)
- **Greeting card, invitation, poster, certificate** — "Write season's greetings to clients", "an invitation to the Christmas party
on 12 Dec at 6 pm": a printable **PDF** from one of four vector templates (A5 card, A5 invitation with When / Where and a QR code for
a link, A4 poster, A4 certificate). In Documents you pick the template and colour and can add your own photo (PNG/JPG — it goes only
into the file, nothing is stored). No AI pictures — the design is vector, logo and your photo.
- **Every document as PDF or Word** — e-mails, notes, meeting points and summaries download on a letterhead (organization name + logo),
with a page footer; e-mails carry To / Subject. Download as text stays.
- **AI helper "Presentation, card, invitation"** in the AI block: the app asks first what to prepare, for whom and in which tone (no
model call), then the assistant continues. The assistant also offers a deck after a project summary or meeting preparation.
- All files are generated **in the browser**; the server never stores them. New document kinds `presentation` and `card` and a `meta`
field on `ai_documents` (migration, no action needed). Private like all documents.
- **Office panel.** The Documents button in the assistant header is now **Office** (briefcase) with the tabs **Documents**,
**Pages** and **PDF** — the separate PDF button is gone (two look-alike icons were confusing). PDF files chosen in the tab
survive switching tabs and closing the panel. **New** next to the filter (in every section, Invoices included) opens a
**New document** form composed by the app (no model call): pick the kind — presentation, greeting card, invitation,
poster, thank-you, e-mail, meeting points, note, project summary, invoice or free text — fill in a few details and
**Prepare** sends the request to the assistant. An empty filter shows the kinds as chips; the AI-block helper opens the
same form. After the assistant reads a map without writing anything, the app adds the chip **Turn this into a
presentation** (the model rarely offered it on its own).
- **Fixed (local models):** when the model asked to open a tool group that was already open, the server re-asked it with the
same prompt and the model repeated the same request until the rounds ran out — the reply came back empty. The model now
gets a one-line note naming the tools it already has and continues; a tool call "written" in the last round without
tools yields a readable sentence instead of an empty bubble. A tool call written as text (`[Tool: name] {…}`, seen with
Qwen-family models with low reasoning) is converted into a real call — only when it stands at the very start of the reply,
and never for tools that change data without a confirmation card.
### Upgrade notes
- Migration `1792001345_ai_documents_kancelar` adds the document kinds `presentation` and `card` and a JSON field `meta` on
`ai_documents` (empty for existing documents). Its timestamp is older than the already applied `1792621271`, `1792928875`
and `1793096888`; PocketBase applies it anyway. Nothing to do.
- New npm dependencies `pptxgenjs` and `docx` (both MIT) are listed in `THIRD-PARTY-LICENSES.md`; the usual
`docker compose up -d --build` picks them up.
---
## v0.95-beta — 2026-11-02
**From an invitation to the calendar: a photo, PDF or e-mail with several dates becomes events on one card**
- **The assistant reads invitations and schedules.** A photo of a lecture schedule on a door, a PDF or a pasted
e-mail with several dates → the new tool `create_events` puts every date on **one card** (weekday, time, place;
*already in the calendar* and *at that time you have …* per line). Repetition ("every Tuesday until January") is
expanded into single days (max 30), excluded days and past dates are left out (a past date that stays is marked
*already passed*), a missing year means the nearest
future date and unclear dates are asked about. Everything or nothing is written after the card is confirmed.
- **Then the app offers a project for the preparation** (no AI call until you click; with two or more events, or
one event from a photo, PDF or pasted e-mail): *Also create a project for the
preparation* builds a project with sign-up, payment, equipment… and the deadlines the invitation states; the event
dates themselves stay in the calendar.
- **Events have a place** (`place`, ≤ 200 characters) — in the event dialog, the calendar, My day, lite, the reminder
text, export/import, REST and MCP. A meeting room to *book* is still a shared calendar.
- **Event series.** Events created together share a `series`; the event dialog offers **Delete the whole series**.
- The image transcription now copies dates, weekdays, times, places and exceptions exactly and keeps tables row by row.
- API: `POST /api/kb/v1/events/batch` (1–30 items, all or nothing, duplicates skipped and reported, one invitation
notification per batch, no time off; a date already in the *same* calendar is skipped, even in a room calendar), `POST /api/kb/v1/events/series/{series}/delete`; MCP `create_events`
(remote `/mcp` and the stdio server in the repository; the published npm package is still 0.63) and `place` on `create_event` / `update_event`. Unknown fields `location`, `venue`,
`where`, `address` now hint at `place` (`location` used to hint at `calendar`).
### Upgrade notes
- Migration `1792621271` adds `events.place` and `events.series` (empty for existing events). Nothing to do.
---
## v0.94-beta — 2026-10-29
**Two map styles: a project to finish, or a process that keeps running**
**Upgrade notes:** migration `1792928875_kind_process` runs automatically on start (adds `process` to `goalmaps.kind`,
`map_kind` to `templates` and the system template *Weekly office operations* — once; a system template you deleted is not
restored); no new variables. Existing maps stay projects and are unchanged; existing recurrence rules keep working. The
"today" a recurring step's next deadline is counted from now follows the instance's `TZ` (default UTC), like scheduled jobs.
- **Project or process** — *New project* offers two tiles: 🎯 **Project** (finish something: % done, archive when everything is done) and 🔁 **Process** (keep something running: steps repeat, nothing gets finished). The style can be switched any time in ⋮ (owner only); steps and deadlines stay, recurring steps keep repeating (in a process they also reset their sub-steps — the server adds or removes that on the switch). A step added with **Add step** is assigned to you. See [Project and process](https://killbottleneck.com/features/project-and-process).
- **Run health instead of %** — a process's apex shows *N on time · M overdue · last run*, the map overview has a **Run health** card (on time, overdue, runs in 30 days, a table of steps), Organization a **Processes** card and tile. Process steps never count as "Not moving".
- **Richer recurrence** — weekdays (e.g. Tue and Thu), day of the month and every N-th (every 2 weeks…). "Every month on the 31st" now really stays on the 31st — a shorter month gets its last day, the next month the 31st again (before, it slid to the 28th for good). A recurring step coming back also clears the "when I'll work on it" plan; in a process its sub-steps start over (new rule action `reset_subtree`). The rhythm shows in words on the card for everyone who can see the map, and in My day.
- **Fixed:** a reminder on a recurring goal was deleted when the goal was ticked off; it now stays and moves with the deadline.
- **Fixed:** the next deadline of a recurring goal is counted from *today* in the instance's time zone (`TZ`, as the scheduled jobs already were), not in UTC — just after midnight in Prague a late step ticked off no longer comes back due "today".
- **API / MCP / assistant** — `kind: "project" | "process"` on `POST /v1/maps` and `create_map`, `kind` on the map list and detail (the org map reads `org` in both), `health` on a process's detail and in the portfolio (a process has `pct`, `done`, `total` = `null`; `counts.projects` still counts all maps, the new `counts.processes` how many are processes), `every` / `days` / `month_day` on `set_deadline` with `advance`. On a public map a visitor who is not signed in sees a process's rhythm and on time / overdue, not when steps were ticked off. A new system template *Weekly office operations*; the former "Process template" label is now **Template with a schedule**.
- **Assistant knows processes** — "make *Stock check* repeat every Friday" sets the step's rhythm the same way as the Recurrence switch (new tool `set_step_rhythm`, one card per step, first due date included), "change it to a project" switches the map style (`set_map_style`, owner only), and reading a map tells the assistant it is a process and how it runs.
- **Projects and processes apart** — the home page has an **All · Projects · Processes** switch (shown once you have a process, remembered), and in Tasks a process is marked *🔁 Process* with *process goal* / *process step* rows and the step's rhythm.
- Migration adds the `process` value to `goalmaps.kind` and `map_kind` to `templates` (no action needed).
- **Fixed:** with some local models (Qwen family, e.g. on llama.cpp-style servers) the assistant occasionally got a tool call
half in the model's own XML form — the rest of the call ended up inside one text field (a project goal reading
`…</goal><parameter name="…">`) and the other fields were lost. The server now cuts the field at the first such tag and
fills in the parameters behind it, only those the tool really has, typed by its schema and never overwriting a value the
model did give. An odd parameter name from the model can no longer break the assistant's reply.
## v0.93-beta — 2026-10-29
**Invoices for anyone in the world: supplier country, any currency (EUR, USD, GBP…), generic tax, free-form bank details, SEPA QR**
**Upgrade notes:** migration `1793096888_clients_mena` runs automatically on start (adds the optional `clients.currency`);
no new variables. Existing supplier profiles and invoices are unchanged (no country = Czech Republic, CZK, SPAYD QR as before).
- **Supplier country.** The supplier profile (shared or own) has a `country` (ISO 3166-1 alpha-2; empty = CZ, so
profiles from v0.86 are unchanged). It presets the currency, tax rate and label, the ID labels (IČO/DIČ · EIN/Tax ID ·
Company No./VAT No.) and the payment block; everything can be overridden (`currency`, `vat_rate`, `tax_label`, `bic`).
Presets (one table for the form and the assistant): CZ, SK, US, GB, DE, AT, PL, NL, FR, IE, CH, CA, AU; any other
country = EU (EUR, VAT, IBAN + BIC). Outside CZ/SK the bank account is **free text**
(routing + account, sort code…) — the server no longer rejects it; a Czech account number or an IBAN is still
recognized and validated. `ico` outside CZ/SK accepts letters and dashes (EIN `12-3456789`, `SC123456`); a tax ID is
required for charging VAT except in the US. New tool fields `country`, `currency`, `tax_label`, `bic` on
`set_invoicing_profile`; the English first-invoice form asks for the country first.
- **Currency per invoice.** `create_invoice` takes an optional `currency`; otherwise the client's currency (new optional
`clients.currency`, set in the clients dialog or by `save_client`), then the supplier's default. The invoice stores
its currency; cards, the assistant's texts, the list and the PDF print it with a symbol (`1 600,00 €`, `$1,600.00`,
`1,600.00 CZK`). `list_invoices` and the Documents list give totals **per currency**, never summed across currencies.
- **QR payment by currency.** SPAYD only for CZK; EUR invoices with an IBAN get an **EPC (SEPA) QR code** (EPC069-12, BIC
optional) read by banks across the EU; other currencies print the payment details without a QR code. DTO: `epc` next
to `spayd`.
- **PDF by country.** Labels follow the invoice language as before; the supplier's country adds the currency symbol and
position, US date order (`10/09/2026`, English invoice only), the ID labels, the tax name ("Sales tax 8.25 %"), no "Not a VAT payer" line
in the US ("Not VAT registered" in GB), date of taxable supply and variable symbol only for CZ/SK (others print a
payment reference), BIC and multi-line bank details in the payment block. **A Czech invoice renders byte-for-text
identically to v0.86** — `tests/faktura-pdf.js` compares against `tests/fixtures/faktura-cz-v086.json`.
- **UI.** Organization settings / My account → invoice details: country select, currency, tax name, BIC, free-form
bank details and a tax disclaimer ("the app does not check the tax rules of your country"). Clients dialog: invoice
currency. Documents → Invoices: labels and dates by the supplier's country, unpaid totals per currency.
- Docs: *Invoices to clients* → new section *Invoicing outside the Czech Republic* (preset table, tax disclaimer).
- A Slovak supplier's domestic account number (`19-8742637541/1200`) yields a **Slovak IBAN** (SK31…) for the EPC QR code;
free-form bank details are capped at 4 lines (address 8). `clients.currency` / `ico` are validated on the server for direct
REST writes too (clients dialog). CAD/AUD print with the code (`1,600.00 CAD`) so totals per currency stay distinguishable.
- Migration `1793096888_clients_mena` (`clients.currency`, text 3). Tests: `faktura-cisla` +25, `faktura-pdf` +17
(US, EUR with decoded EPC QR, GB, Czech supplier in EUR, v0.86 fixture), `ai-faktury` +2 sections (client currency →
EUR invoice, explicit USD, per-currency totals, invalid currency, US member profile, routes).
## v0.92-beta — 2026-10-28
**Maintenance release for the hosted service. Nothing changes for self-hosted servers.**
**Upgrade notes:** none — no migrations, no new variables for your own server.
## v0.91-beta — 2026-10-28
**Fix: pages scroll again with the mouse wheel. Since v0.90-beta, Projects, Tasks, Calendar and every other page with the
top bar could not be scrolled with the wheel or the scrollbar, so items further down were out of reach.**
- **Scrolling restored.** The scroll lock meant for an *open* dialog from the avatar menu (My account, Appearance, …) was
applied permanently, because those dialogs are mounted even while closed. The lock now applies only while such a dialog
is actually open and is released when it closes.
- The pinned top bar behaves as intended: like a frozen header row — the bar stays at the top and the content below it
scrolls with the wheel.
**Upgrade notes:** none — no migrations, no new variables.
## v0.90-beta — 2026-10-27
**The top bar with the bell and your avatar now stays put everywhere: the assistant panel, the start-up aquarium, the left
panels and the dialogs from the avatar menu all start below it, and Organization settings has the bar too.**
- **Sticky top bar across the whole window.** Projects, Tasks, Calendar, Organization, Archive, Notifications and
Organization settings keep the bar at the top while you scroll. It spans the full window width like the map editor's
bar, so the avatar sits at the right edge on every page.
- **Assistant below the bar.** The side panel, the aquarium shown next to the chat, the Documents panel, the idea buffer
and the time log all begin at the bottom edge of the bar — the bell and the avatar are never covered. This replaces two
earlier behaviours (the panel over the editor bar on wide screens, the aquarium over the bar). Full-screen relax with
the PIN lock still covers everything, as before.
- **Dialogs from the avatar menu** (My account, Appearance, Invite, Time worked, API keys, AI agents, Connect agent,
Report a problem, About) open below the bar and the bar stays clickable: opening notifications keeps the dialog,
clicking the dark veil closes it. Only one of these dialogs is open at a time.
- **Organization settings** has the application bar (with the back arrow in it) like every other page.
- **Editor banners** (“map changed elsewhere”, “project is archived”) move the bar and everything below it together.
- **Phone:** the assistant panel and Documents start below the two-row bar. Lite mode is unchanged.
- Hosted instances: assistant Pages are generated on the customer's own model too (no quota), when one is set.
**Upgrade notes:** none — no migrations, no new variables.
## v0.89-beta — 2026-10-25
**Own model in the cloud: a hosted instance can run the assistant on the customer's own OpenAI-compatible service (home GPU behind a tunnel, own server or a provider of their choice); AI settings of the operator are now locked server-side**
- **Organization settings → Own model (hosted instances only).** The administrator enters the address of an
OpenAI-compatible API (usually ending in `/v1`), the model name and, if needed, a key, tests the connection and switches
the assistant over. Turns on the own model **do not count towards the weekly quota** (logged with `zdroj=vlastni`,
0 credits); the hourly per-user brake stays. Photos for the assistant and voice transcription keep running on the
operator's AI and keep using the quota. When the own model is down the assistant says so (`502`, `code: ai_vlastni`)
and never silently falls back to the operator's AI. New routes `GET/POST /api/kb/ai-vlastni` and
`POST /api/kb/ai-vlastni/test` (administrator only, `404` on self-hosted instances); `/api/kb/config` carries
`chat_vlastni`. Guide: [Own model in the cloud](https://killbottleneck.com/tutorials/own-model-in-cloud).
- **Private-address guard now judges IPv6 by value, not by spelling** (self-hosted AI settings too): `[::ffff:7f00:1]`,
`[0:0:0:0:0:0:0:1]`, `[::1%25lo]`, NAT64 `64:ff9b::/96`, 6to4 and link-local forms are rejected like `127.0.0.1`
(they used to pass). With an own model, turns carrying a photo or a voice note still use the weekly quota (vision and
transcription run on the operator's AI); only the text calls to the customer's model are free. Extra fields of the own
model cannot override `messages`, `tools`, `model` or `stream`. A stored key that can no longer be decrypted (changed
instance key) stops the assistant with a clear message instead of calling the model without a key.
- **The key never leaves the server.** It is stored encrypted (AES, `KB_ENCRYPTION_KEY` of the instance), `GET` returns
only `token_set`, and it is sent only to the address it was saved for: a connection test against another origin goes
without it, and changing the address to another origin without a new key deletes the old one. Private, LAN and
Tailscale (`100.x`) addresses are refused on hosted instances — the model needs a public address (tunnel).
- **Security (hosted instances):** `GET/POST /api/kb/ai-settings` and `POST /api/kb/ai-test` now return `403` on a hosted
instance — until now only the UI hid them, so a customer administrator could redirect the operator's AI through the
REST API. `POST /api/kb/ai-test` adds the stored token only when the tested address has the same origin as the saved
one (before, the stored token went to whatever address the request named), and the first save of AI settings copies
the token from `.env` only when the address stays at the same service. The chat re-checks a database-stored AI address
against the private-network guard right before every call, like `/advisor` already did.
- `llm.js` sends `Authorization` only when a token is set (an own `llama-server` without `--api-key` got an empty
`Bearer ` header before).
**Upgrade notes:** migrations add the `ai_vlastni` collection and `ai_chat_log.zdroj`. Hosted instances need
`KB_ENCRYPTION_KEY` (exactly 32 characters; `tenant-add.sh` generates it once and keeps it) — without it the Own model
section says so and stays off. Self-hosted instances are not affected: they keep the AI tile in Administration and the
new section does not appear. Hosted instances that had saved `ai_settings` through the API lose that path (`403`); the
assistant on hosted instances has always read `KB_CHAT_*` from the environment, so nothing changes for them.
## v0.88-beta — 2026-10-25
**Pages: the assistant (or a connected agent) writes a single-file HTML page — presentation, overview, small app — and you share it by a secret link or with the team**
- **New tab Pages** in the panel next to the chat (beside Documents). Tell the assistant *"make me a presentation of
the Q4 results for management"*: the assistant reads the project with its tools and hands over a brief; **the app
then writes the whole HTML in a separate model call after the turn** (`create_page` only queues it — large HTML never
travels through tool arguments or the chat history; the page writer runs with `KB_PAGE_NUM_PREDICT` output tokens and
the result arrives through realtime, so a Cloudflare-cut response does not lose it). Rewrites via `update_page` keep the
previous version (**Restore**); a failed write offers **Try again**. Rename by hand only; download as `.html`.
- **Templates, so the feature is not hidden behind the right wording:** **New page** in the Pages tab (and *Page from a
template* among the AI helpers and the chips under the chat) opens a form — template (presentation for management, team
overview, customer summary, timeline, status report, checklist, custom brief), project, look, notes — and the app composes
the message to the assistant.
- **Visuals by default:** the outline assigns every section a visual (stat cards, progress bars, Gantt, timeline,
donut, traffic light, table, checklist) and an icon, the skeleton ships a small CSS toolkit for them and each
section draws its chart as inline SVG / CSS from the given facts — no external images; the templates ask for
them explicitly (timeline = Gantt with a today marker, status report = traffic-light dots + cards).
- **You can watch the page being written:** the writer builds it **section by section** (outline → look and skeleton →
each section) and after every step stores the partial HTML; the panel shows the list of steps with the running one,
the preview grows, and a **Code** toggle shows the HTML as it comes in. (PocketBase calls the model synchronously, so
token streaming is not possible — this is the honest substitute; smaller calls also suit local models better.) A
rewrite of a page built this way touches only the changed sections (keep / rewrite / new / delete), a page without
section markers (uploaded by an agent) is rewritten as a whole.
- **Sharing like a project:** **Share by link** turns on a secret `/p/<32-char token>` link that anyone opens without an
account (**New link** invalidates the old one, **Turn the link off** stops it at once, 60 opens/min/IP); **People** lets
you pick **named members of the organization** (as in the map's share dialog; `shared_with`), or the **whole team**
— the other person finds the page in the **Shared with me** tab (never the token). The assistant does the same with
`share_page user_on / user_off` + `email`, MCP and `/v1/pages/{id}/share` too. The link carries the address the
request came in on (like a project link: LAN → LAN address, cloud → the instance domain); `KB_PUBLIC_URL` is used only
when the request has no usable host (MCP over 127.0.0.1, cron). Pages are in
**Download all my data** (without the token). Public pages keep working after a trial expires (like public maps);
writes are 402.
- **Sandbox, not sanitizing:** `/p/…` is served with `Content-Security-Policy: sandbox …` (no `allow-same-origin`),
`nosniff`, `noindex`, `no-referrer`, `no-store`; the in-app preview and the `/stranka/:id` window use an `<iframe sandbox>`
without `allow-same-origin` — the page runs as an opaque origin and cannot read the app's sign-in (localStorage),
cookies or API. Opening in a new tab goes through `/stranka/:id`, never a `blob:` URL (which would inherit the app origin).
- **MCP:** six new tools for connected agents (`list_pages`, `get_page`, `create_page`, `update_page`, `share_page`,
`delete_page`; HTTP and stdio, 32 tools in total) backed by `GET/POST/DELETE /api/kb/v1/pages…`; there the agent writes the
whole HTML itself. `/agent-setup.md` mentions pages.
- **Limits and switches:** `KB_PAGES=0` turns the feature off (tools, tab, `/p/` links), `KB_PAGES_MAX` (100; Cloud Lite 20),
`KB_PAGE_KB` (300), `KB_PAGE_NUM_PREDICT` (12000; Lite 8000), `KB_PAGE_TIMEOUT` (280 s), `KB_AI_PAGE_VAHA` (3 in the hourly
AI limit). A page is text, not an upload — it works with `KB_FILES_MB=0`; V1 has no attachments.
- **App start with the aquarium and the assistant open:** on a wide screen the app opens the assistant panel and starts
the aquarium beside it (relax without a countdown), once per browser tab; closing it and reloading does not bring it
back. `KB_START_RELAX=0` turns it off. Every relax scene has its own **Close** button next to the light setting.
- **"My map" is gone from the top bar** (still reachable from My day and `/my-map`).
- **After review (security, quality, architecture):** a shared page keeps serving its **last finished version while it is
being rewritten** (as a project does); pages left in "being written" by a restart turn into a retryable error; the
rewrite sends the model the **skeleton and text excerpts** of the sections, never the whole HTML (a page without section
markers over 100 kB is refused with advice to edit it by hand); one write is capped at **15 minutes** and the size limit
is checked after every section; at most **2 pages per user** are written at once; the public page adds
`form-action 'none'; connect-src 'none'; frame-ancestors 'none'` to its CSP and the in-app preview carries the same
`form-action`/`connect-src` rule, so a page uploaded by an agent cannot submit forms or send data anywhere; **Restore** is
unavailable while a page is being written.
- **Fixed:** the service worker cached **every** navigation response as the offline shell (`/`), so visiting
`/p/<token>` or `/agent-setup.md` could replace the app shell with foreign HTML — those paths are no longer intercepted.
## v0.87-beta — 2026-10-24
**Come-back e-mails: whoever has not signed in for 3, 7, 14 or 30 days gets an overview of their open work and the smallest steps back**
- **Come-back e-mail ladder** — after 3, 7, 14 and 30 days without opening the app an e-mail arrives with what is waiting, what is overdue, what is holding up others, and three smallest steps for today. The text is written by the AI assistant from the same data as the morning summary; without AI the e-mail still goes out with a plain list of the most important items. One e-mail per step, only the highest reached step, silence after the last one until the next sign-in; signing in resets the ladder.
- **Unsubscribe in one click** — every such e-mail carries an unsubscribe link (also `List-Unsubscribe` for mail clients) and the same switch sits in Notification settings. Not sent in the "Nothing" e-mail mode, after an expired trial or without SMTP.
- New environment variables `KB_NAVRAT_MAILY`, `KB_NAVRAT_DNY`, `KB_NAVRAT_HODINA`, `KB_NAVRAT_MAX_DENNE` (see [Environment](https://killbottleneck.com/reference/environment)); a new `navrat` field on `users` (migration, no action needed).
**Every e-mail subject starts with the organization name**
- **Subjects like `Tengo: Welcome to killBottleneck`** — invitations, welcome mail, password reset, verification, one-time code, notifications, the daily digest, project-share invitations and bug reports all start with the organization name (the word from the instance address; a self-hosted instance on its own domain uses the name from Organization settings, without one the subject is unchanged). The former "— organization tengo" suffix on invitations and the welcome mail is gone, the name now leads.
- **The organization name also heads the body** — right under the logo of every e-mail, large, with killBottleneck underneath, so a recipient who does not remember the app still recognizes their own organization. Without a known organization the header is unchanged.
- **"Login from a new location" in the recipient's language** — the alert about a sign-in from a new device was the only system e-mail still sent in English and without the organization name. It now uses the same look, shows when, from which address and browser, and its button leads to a password change (a new password signs out every other device).
## v0.86-beta — 2026-10-24
**Invoices to clients from the assistant — PDF with a QR payment code; client address book with billing details; supplier profile**
- **"Issue an invoice to Karel for 2 hours of consulting at 800."** New assistant tool group `faktury`: `list_clients`,
`save_client` (also from a business-card photo transcript — the card shows exactly what will be saved, nothing is
invented), `set_invoicing_profile` (scope `org` = shared, administrator; `me` = own), `create_invoice`, `list_invoices`. The invoice card lists every
item and the totals **computed by the server** (the model never computes amounts); after confirmation the app assigns
the number (one series per year — shared by the organization or the user's own, `20260001`, no gaps — unique index + transaction, two parallel
confirmations get two different numbers) and the card offers **Download PDF** and **Open in Documents**.
- **PDF is created in the browser** (pdf-lib + Liberation fonts, lazily loaded): header with the organization logo,
supplier and customer, dates (plus date of taxable supply for VAT payers), items table over several pages, totals,
note, payment block with account, IBAN, variable symbol and a **QR payment code (SPAYD)** — the IBAN is derived by
the server from a Czech account number (check digits verified, mod 11 and mod 97). The PDF is never stored; the
invoice keeps snapshots of the supplier and customer, so it can be regenerated identically.
- **Supplier details, shared or your own.** The organization's shared details are in Organization settings →
*Issuing invoices* (administrator). **Every user can also set their own details** under *My account → My invoice
details*: they take precedence over the shared ones for that user's invoices and give them **their own number
series** (a team of freelancers in one instance, each invoicing for themselves). When the applicable details are
missing and someone asks for the first invoice, **the app shows a form** (no model call; an administrator first
chooses "whole organization" or "only me") and the assistant saves the answers. VAT: non-payer, or payer with one
rate per invoice; CZK only.
- **Clients** (the existing address book from time tracking) gained optional billing fields: address, company ID,
VAT ID, e-mail — editable in the clients dialog (*Billing details*) or by the assistant. Ambiguous first names
("Karel" = two clients) make the assistant ask which one.
- **Documents → Invoices** has a **filter by month issued** and **Download all (ZIP)** — the PDFs of the listed invoices in
one file for the accountant (fflate, MIT; lazily loaded). *My account → My invoice details* links there; the "Download all
my data" export says it does not contain invoices. Cards "Create the client" / "Save the supplier details" link to the
clients dialog / Organization settings / My account.
- **Documents → Invoices**: list of issued invoices (own; administrators see all), detail with items and payment
details, *Download PDF*, *Paid* toggle. Invoices are immutable — a correction is a new invoice. The app says plainly
that it is not the legal archive of tax documents: keep the PDF.
- Routes `GET/POST /api/kb/invoicing-profile` (admin), `GET/POST /api/kb/invoicing-profile/me` (own details; `{smazat: true}` removes them), `GET /api/kb/chat/faktury`, `GET /api/kb/chat/faktura/{id}`,
`POST /api/kb/chat/faktura/zaplaceno`; collection `invoices` (read: owner or admin; writes only by the server),
migrations for `clients` fields and `instance_settings.invoicing`.
- **Assistant help answers "how do I…" without doing it.** The help tool's result now opens with a reminder that the user
asks HOW something is done, so the model explains the steps and only offers to do it (a `suggest_next` chip) instead of
changing the setting itself — DeepSeek used to switch the theme right after "How do I switch to the dark theme?".
- Legal texts (DPA 2.2, 6.1, 7.1; privacy policy 3.3, 4.1; terms 2.2): data categories extended by client billing
data and issued invoices; AKI.IO GmbH listed as the AI sub-processor (EU, zero retention) instead of the outdated
"AI on the operator's hardware in the Czech Republic" wording.
## v0.85-beta — 2026-10-19
**The assistant stays visible when the AI is down or not connected: history, Documents and relax keep working**
- **The assistant no longer disappears when the AI does not respond.** The header button (and the one in the map
toolbar) and the side panel stay; a bar at the top of the panel says the AI is temporarily unavailable, offers **Try
again** and lists recent conversations so you can open them. Documents and relax (the word "relax" or the 🐟 chip) keep
working. Typing, voice messages and photos are paused — a message you type stays in the field. As soon as the AI
answers again the bar disappears by itself; a failed message also re-checks the AI right away instead of waiting.
- **Installations without AI now show the assistant too.** The bar explains what the assistant will do once AI is
connected; admins get **Connect AI**, which opens *Organization admin → AI*, members see who can connect it. There is
no edge tab or phone bar until the AI is connected, only the button in the header or the map toolbar. If the AI is
configured but does not respond, the bar says so and admins get a link to check the AI settings.
- Buttons elsewhere that need the AI (sorting the idea buffer, the morning and night planning offers, the event dialog)
hide while the AI is down, as before.
- **Picking a look:** the area around the dialog is only slightly dimmed, so you see the new look right behind it.
- **Relax next to the chat** no longer covers Documents, and links from the assistant's cards ("Show in map", "Open
project"…) end a relax that is not locked, so the page they open is not hidden under the aquarium.
Upgrade notes: no configuration changes.
---
## v0.84-beta — 2026-10-18
**The assistant shows where things go, gives steps icons automatically and structures its answers; Tasks mark sub-steps with `└`**
- **The assistant's confirmation card shows where things go.** Putting an idea into a project, moving a step or adding
steps under a step no longer reads as one long sentence: the card draws the path — the idea buffer (or the step's
current place) on top, an arrow, then the project, the parent steps and the new or moved row highlighted. The sentence
above it only names the action and the project.
- **The assistant's answers are easier to scan.** A part of the answer that starts with a label ("Done:", "Blocked:",
"On fire:", "Recommendation:" …) is shown as its own block with an icon and a coloured edge instead of one long
paragraph.
- **Tasks: sub-steps carry a `└` mark** in front of the name, so a row visibly belongs under the one above it (rows with
their own sub-steps keep the expand arrow). The target circle in front of every step name is gone; only the step's own
icon shows when you set one.
- **The assistant gives steps and projects icons automatically.** When it creates a project, adds steps or moves an idea
from the buffer, every step gets a fitting emoji (and the project its own); the confirmation card shows them. Ask for
icons on an existing project ("add icons everywhere") and the assistant proposes them for all steps in one card. REST
API v1 and MCP: nodes accept and return `icon`; the map text the agents read shows icons in front of step names.
Upgrade notes: no configuration changes. REST API v1 and MCP nodes carry an optional `icon` field (up to 16 characters).
---
## v0.83-beta — 2026-10-16
**Team calendar: shared calendars (company events, meeting room, time off), the calendar as its own page, the assistant handles it all — never per user**
- **Calendar is its own page** (`/kalendar`) with a **Calendar** item in the main navigation for every role; the old
`/tasks?view=calendar` link redirects there with its parameters. The Tasks page no longer has the Calendar tab.
- **Security (panel 15 Oct)**: the read rules of `calendars` and `events` now require a signed-in user (migration
`1792080000`; the previous rules let an anonymous request read calendars and events, the personal-event part of that
predates this release), and they match the routes: a rejected request is visible only to the requester, the person
concerned and admins/managers, events of archived calendars are not returned. Approval cannot be bypassed by editing:
moving an event into an absence calendar with approval, or changing the dates/person of a decided request, creates a
new request (the decision is cleared, admins and managers who see the calendar are notified again). Only an admin or
manager can enter time off **for a colleague** (`err.absenceForOtherAdminOnly`); `decision_note` is returned only to the
requester, the person concerned and admins/managers (also in the export); leaving an event you were not invited to is
refused (`err.eventNotParticipant`); the event list cap is 3000 with a `truncated` flag; import matches calendar names
exactly.
- **Default calendars**: a new instance starts with *Company* (`shared`) and *Time off* (`absence`) in the first
administrator's language (`KB_UVODNI_MAPA=0` switches this off, as with the starter projects).
- **Shared calendars** — calendars that are not a person: *company events* (`shared`), *room / resource*
(`resource`) and *time off* (`absence`). An administrator creates them in Organization settings → Shared
calendars (one-click presets: Company, Time off, Meeting room), an admin or manager also through the assistant
or the API, also right from the calendar (**+ New** in the panel, **+** → *Shared calendar*). **Visibility like a
project**: the whole team, or selected members only (plus the creator — an admin who is not listed does not see it, like a private project; `visibility`, `members`).
Every member adds entries unless `write_access` is `admins`. A pencil next to the calendar in the side panel edits it
(name, who sees it, who adds entries, approval), archives it or **deletes it with all its events** (confirmation
shows the count). Side panel: a row per calendar with its colour; hidden ones are remembered.
- **Events have an end** (`end_time`, `end_day`): a multi-day event is a bar across the days (moved in the detail,
not by dragging; moving the start keeps the length). **Room bookings never overlap**: the server refuses a taken
slot with **409** `{ collision, next_free }`, the dialog and the assistant offer the next free slot that day;
a booking without an end takes one hour.
- **Time off**: all-day with a From–To range, an admin or manager can enter it for a colleague; the **Who is away**
band in Week/Day, the **Away today** line in My day and lite, `GET /api/kb/away`. **Approval is optional and off
by default**: when switched on for the calendar, a new entry waits (`requested`), admins and managers get a
*Time-off request* notification, decide in the detail (Approve / Decline + note) or with the assistant — the
**time-off approval mode** gathers who else is away, colliding company events and the requester's due work and
gives a recommendation; the requester gets a *Time-off decision* notification.
- **Assistant**: `create_event` / `update_event` take `calendar`, `end_day`, `end_time`, `person`; new
`list_calendars`, `who_is_away`, `create_calendar` (admin/manager), `decide_absence`; the events tool group
also opens on vacation / room / time-off wording. Cards: "Book "Meeting room" 15 Oct 10:00–11:00", "Record time
off: jana@… 20–24 Oct".
- **Lite**: read-only **Calendar** page for the next 7 days (own events, company events, bookings, who is away) in
the bottom navigation.
- **API / MCP**: `GET/POST /api/kb/v1/calendars`, `POST /api/kb/v1/calendars/{id}`, `GET /api/kb/v1/away`,
`POST /api/kb/v1/events/{id}/decide`; `/v1/events` takes `calendar`, `end_day`, `end_time` and `?calendar=`;
foreign fields `room`, `location`, `vacation`, `end`… get a hint. MCP has 26 tools (was 20): `update_event`,
`delete_event`, `list_calendars`, `create_calendar`, `delete_calendar`, `who_is_away`. Export carries `calendars` and shared-calendar
events; `/import-all` imports calendars (admin/manager) and the importer's own events.
- Permissions are enforced on the server (`can_edit` in the event DTO): the author; in a shared calendar also
an admin/manager; time off also the person it concerns. Personal events stay private even from administrators.
- **Upgrade notes:** migration `1792001003` adds the `calendars` collection, new event fields and two notification
types; the down-migration **deletes** shared calendars and the new fields. Reminders on shared events go to the
author and invitees only (not the whole team).
## v0.82-beta — 2026-10-15
**Relax: a guarded break with an aquarium, a beehive or an ant colony, right next to the assistant**
- **The fish button in the assistant's header swaps the map for the scene** and back; **type `relax`** to the
assistant, pick the chip in an empty chat, or let the assistant offer a break when you write that you are
tired — a calm **aquarium**, **beehive** or **ant colony** scene starts (scene picker on the right edge). A countdown
is optional: `relax 15`, the **5 · 15 · 30 min** presets next to "Relax running", or the assistant's offer; **+5** extends it. Nothing goes to the AI model for the word itself; the assistant's
`start_relax` tool only appears when the conversation is about tiredness or a break, and only after you agreed.
- **Two layouts:** full screen, or **next to the chat** (the scene takes the whole left side, the chat stays on
the right and keeps working); enlarging temporarily from there returns next to the chat with ✕ or Esc. On phones
always full screen. The toolbar fades when the mouse rests, the countdown stays readable.
- **Lock the screen with a 4-digit PIN** and step away: the lock survives a page reload and covers the app
before anything else is shown; five wrong PINs pause input for 30 s. It is a curtain for passers-by, not
security — the dialog says so and points to the system lock (Win+L).
- **When the countdown ends** a soft chime plays, the scene dims and both the scene and the chat offer
**Back to work** / **5 more min** / **Keep relaxing**; a locked screen stays locked until the PIN is entered.
- **The first time you open the assistant each day** the aquarium starts next to the chat by itself (once a day).
The assistant also offers a break **after ~50 minutes of continuous work** (at most once an hour) and adds a
**Relax 5 min** chip after a bigger turn (a guide or three or more tools). The scenes are self-contained
static files (`/relax/*.html`, English copies for the English UI), loaded only while a break runs; they are never
cached by the offline service worker.
Upgrade notes: no configuration changes.
---
## v0.81-beta — 2026-10-15
**killBottleneck for Windows, and a server outage no longer signs you out**
- **New: killBottleneck for Windows** — a separate desktop app (own repository
[tengolabs/killbottleneck-desktop](https://github.com/tengolabs/killbottleneck-desktop), own version numbers) that opens
your organization's server in its own window: Windows notifications also with the window closed, unread count by the
clock, starts with Windows, updates itself. The window shows the server's own web, so new features arrive with the
server. This release adds the web side: on the sign-in page the app shows the organization's address with **Change
organization**, unsaved map changes are saved before the app quits or switches organizations, clicking a Windows
notification opens its target and marks it read, Google Drive picking is hidden in the app (it cannot work outside a
browser), and About shows the app version. Older servers work with the app too (notifications included), just without
these extras.
- **A server outage no longer signs you out.** The client used to clear the stored sign-in on any error while refreshing
it (502 during a deploy, no network yet after waking up); now only when the server rejects it (401/403).
- **After a deploy, an open page reloads itself** when a code chunk it needs is gone (`vite:preloadError`), instead of
failing to load a dialog or page.
- The assistant's built-in help index was regenerated from the corrected documentation.
Upgrade notes: no configuration changes.
---
## v0.80-beta — 2026-10-14
**Sharing with an address that has no account: invite to the organization right from the Share dialog, or a request to the administrators**
- **Share dialog asks before it sends anything** when the address has no account and self sign-up is closed
(registration key, cloud). Someone allowed to invite (administrator, manager, structure manager) gets
**Invite to the organization and share**: the account is created as a Member (`invited_by` set), a password-setup
e-mail goes out (without SMTP the dialog shows the temporary password once) and the project is shared — one e-mail,
which names the project. The second option is **turn on the public link** for a look only. Nothing is created without
that confirmation; the quiet share from the owner selector and the assistant's `share_map` never create accounts.
- **A member who may not invite** shares as before; the recipient's e-mail now says the administrators were asked
(instead of "reply to this e-mail"), and the instance administrators and structure managers get an in-app
**request** in the bell — one per address regardless of the number of projects, and only when the invitation itself
passed the daily caps. Clicking it opens Organization settings with the invite dialog pre-filled.
- `POST /api/kb/share` has a new read-only action `check` (`registered`, `external`, `registration_open`,
`can_invite_to_org`, `needs_confirm`) and `share` takes `invite_to_org: true`; the response carries `invite`,
`registration_open`, `can_invite_to_org`, `admin_asked`. `POST /invite` and the share path create accounts through
one shared function.
- **Fixed:** the notification DTO dropped `extra`, so notifications that carry a payload (the "offer the assistant" action
since v0.76, the new request) did not open their target.
- The "new version" notice in the bell is sent only for an exact release tag, not for interim builds between releases.
## v0.79-beta — 2026-10-13
**Connect your AI agent with one line; OAuth fixes for self-hosted instances and editor clients**
- **"Connect your AI agent" after the first sign-in.** The full app shows a dialog with one line to paste into
Claude Code, Codex or Cursor (`Set up killBottleneck for me: fetch and follow this guide
https://YOUR-DOMAIN/agent-setup.md`; the Czech app shows the Czech sentence). The agent fetches the guide, registers the instance's
remote MCP server (`/mcp`) by itself and asks the user to sign in in the browser (OAuth) — no password or key goes
into the agent. The dialog shows once per user (also to existing users after this update), then stays in the user
menu as **Connect AI agent**. Not shown in the simplified view. `KB_AGENT_SETUP_ASK=0` turns it off; `KB_PURPOSE_ASK=0`
turns it off too.
- **Every instance serves its own agent guide at `/agent-setup.md`** (public, English, instance address and version
filled in): rules for the agent, steps for Claude Code / Codex / Cursor / other clients, the user's sign-in step,
verification via `list_maps`, the API-key fallback and a ✓/⚠/✗ report format. Over plain `http://` on a non-localhost
address the guide sends the agent straight to the API key, because MCP clients only send OAuth credentials to HTTPS
or localhost. Also linked from the OAuth resource metadata (`resource_documentation`).
- **OAuth works on a fresh self-host.** The instance address in the OAuth metadata, `WWW-Authenticate` and the guide
no longer falls back to PocketBase's default `http://localhost:8090`: a localhost value counts as unset and the address
the request came in on (`Host`, `X-Forwarded-Host`, `X-Forwarded-Proto`) is used. Set `KB_PUBLIC_URL` behind a proxy
that does not send `X-Forwarded-Proto`.
- **OAuth for Claude Code re-authentication and editor clients.** Loopback redirect URIs are accepted on any port
(RFC 8252 §7.3 — Claude Code picks a random one; host and path still have to match exactly), and clients may register
`cursor://`, `cursor-insiders://`, `vscode://`, `vscode-insiders://` and `windsurf://` redirect URIs. The consent page
takes the redirect verdict from the server and shows the full redirect URI for editor schemes. Token responses carry
`Cache-Control: no-store`; `scopes_supported` is in the resource metadata.
- **Fixed:** `?next=` after registration now works (the first user of a new instance coming from an OAuth consent is
sent back to it), and `next=https://YOUR-DOMAIN//evil.com` no longer redirects off-site after sign-in.
- Docs: MCP reference has a new "Let your agent set it up" section; `KB_AGENT_SETUP_ASK` and `KB_PUBLIC_URL` notes in
the environment reference.
**Upgrade notes:** a migration adds `users.agent_setup_seen` (no backfill — every existing user sees the dialog once;
set `KB_AGENT_SETUP_ASK=0` before updating if you do not want that). The MCP npm package did not change.
---
## v0.78-beta — 2026-10-13
**Maintenance release**
- **Self-host admins: optional second assistant model.** `KB_CHAT_ROZJEZD=N` hands a conversation from the primary
model to a second one (`KB_CHAT2_*`) from the N-th turn on, when it has a free slot; text drafts (e-mails, letters,
talking points) always stay with the primary model, writes can too (`KB_CHAT_ROZJEZD_CO=cteni`), tool groups for the
second model come from the conversation history (`KB_CHAT2_TOOLS=historie`), the client pre-warms it after every reply
(`POST /api/kb/chat/predehrat`), and a failing second model falls back to the primary within the same turn.
`KB_CHAT2_STIN=1` returns a shadow answer for comparison. Without `KB_CHAT_ROZJEZD` nothing changes. New test `ai-rozjezd`.
- Test harness waits for the container's network to settle before opening the browser (random `ERR_NETWORK_CHANGED`
on the first page of a click suite).
- No migration, nothing to configure.
---
## v0.77-beta — 2026-10-12
**From the map to the tasks of this project only — and back (the phone got lost in the task list)**
- **New ☑ "Tasks of this project" icon in the map's left rail** (below the progress dashboard): opens the task list
narrowed to this project (`/tasks?map=<id>`). The same entry stays in the toolbar's ⋮ menu; on the phone nobody
found it there, and from the task list there was no way back — the header arrow led to the home page and the list
then showed every project. The "Report a bug" rail button moved one slot down.
- **Project strip on the Tasks page**: "Project tasks: *name*" with a **Back to map** button and a × that widens the
list to all projects again; the header's back arrow also returns to that map.
- The project filter is written to the address (`?map=`) when chosen in the dropdown too, so a reload or a shared link
stays on the same project.
- New click-test `ui-ukoly-projektu` (phone + desktop). No migration, nothing to configure.
---
## v0.76-beta — 2026-10-06
**The assistant shows what it is doing while you wait; smarter search across projects and documents; risks highlighted in the map; rule action "offer the assistant"**
- **Search understands word forms and ranks results.** `search_projects` (the assistant's search across active and
archived projects) now uses the same mechanics as the help lookup: query words are matched by stem in both directions
("fakturami" finds "Faktura"), stop words are ignored, rare words weigh more than common ones, and a multi-word
query does not have to match completely — results are ranked by fit (project title > step title > owner >
description, adjacent word pairs score extra). Substring matching stays as a fallback, so old queries behave the same.
- **Search covers your Documents too.** Drafts, meeting notes and prep notes written by the assistant are searched
alongside projects (own documents only) and listed after the projects with a short excerpt.
- The assistant is told it can pass what the user said as the query (a few key words in any form) instead of
guessing stems.
- **Risks right in the map.** New assistant tool `map_risks`: for one project it lists the steps that are overdue,
blocking open sub-steps, waiting for sub-steps or idle 14+ days (same definitions as My day / Organization, facts
with day counts, no scoring) and attaches a card **Highlight in the map** — the steps get a temporary colour in the
editor with a small bar (counts + Clear). Browser-only, nothing is written to the map; no new toolbar button.
- **New rule action "offer the assistant"** (`offer_assistant`, `mode` + `to` as in `notify`): the rule sends a
notification whose link opens the assistant right in the chosen package (Review, After the meeting, Meeting prep,
Weekly review, Sort notes, Morning briefing, Evening planning) for the goal that fired it — e.g. "Client meeting"
done → offer After the meeting; Friday 16:00 → offer the Weekly review. Nothing runs until the person clicks, so it
costs no AI turn; writes still go through a confirmation card. Notifications gained a small `extra` field
(migration), the map deep-link accepts `?asistent=<mode>`.
- Hybrid light/heavy setups: questions about a project's state ("how is X going", "what's overdue", "where are we
stuck") now go straight to the main model — the light model read the map correctly but invented finished steps in
its summary. Reply cleanup also drops a bare `{"suggestions":[...]}` line (turned into chips) and a leaked
"Need to respond." opener.
- While the assistant works on your message, the waiting bubble no longer shows only "Thinking…": it names what
the assistant has just read (e.g. "Read: project map „Workshop“") and whether it is preparing a proposal for
you to confirm. The text is composed by the app from the tool names — nothing extra is sent to the model.
- Self-hosted instances with a local model benefit most (the first turn after a model load can take a minute).
- New collection `ai_chat_prubeh` (one small record per user, written only by the server, read via realtime);
the migration creates it on upgrade. Nothing to configure.
**The assistant answers how-to questions from the user guide**
- **`help` tool for the assistant**: when the user asks how something is done in the app or where a button or
setting is, the assistant looks it up in the same user guide people read on the web (`docs/cs/jak-na-to`,
`docs/cs/funkce`, `docs/tutorials`, `docs/features`) and answers from the returned excerpt with the exact
button names; the reply carries **More in the guide:** links to the matching pages (`.cz` for Czech users,
`.com` for English). No separate "cheat sheet" for the AI: the guide stays the single source. The index
(`pb_hooks/napoveda-index.js`, one entry per heading H2–H4 with its anchor) is generated by
`docs/scripts/gen-napoveda.mjs`; `--check` runs in `npm run kontrola`, `nasadit.sh` and the test suite, so
the index cannot drift from the docs. Token cost: the tool schema (164 tokens) is offered only for how-to
questions (own keyword group), one lookup returns at most 2 500 characters (~440–830 tokens); the full Czech
guide would be 64–75k tokens per call, which is why it never goes into the prompt. Checked live on the production model
set (DeepSeek V4.1 + gpt-oss): the model rewrites the query into settings jargon, so the lookup also searches with the
user's own sentence (which outweighs the rewrite); the light model invented a page filter, so the tool has no `page`
parameter; and because the light model once *switched* the theme when asked *how* to switch it, how-to questions go
straight to the main model in the hybrid and the prompt says a how-to question is not an instruction to do it.
## v0.75-beta — 2026-10-05
**Assistant sends the model less data; AI quota counted in real turns**
- **Assistant: fewer tokens per call, prompt cache holds.** The map list in the system prompt is ordered by project
number instead of last change, so a confirmed write no longer reshuffles it and the provider's prompt cache keeps
the whole history. Tool results have caps (`get_settings` lists 30 members, `list_people` 150, `get_memory` 5
projects × 1 500 chars). Messages longer than 5 000 chars and tool arguments over 3 000 chars (a long draft, a
project outline) are shortened in the history once they are older than 3 turns; the guided modes (sorting, briefings,
new project) keep the full history. Tool groups open on narrower keywords — "objednání" no longer opens meetings,
"osvětlení" the look, a bare "uprav/změň" the documents, a bare "upozornění" the rules. New measurement suite
`ai-tokeny.js` and log columns `skupiny`, `prepis_ms`.
- **Assistant: slimmer system prompt, tools can be opened by the model.** The system prompt keeps only the general
rules (about 40 % shorter); the instructions for rules, projects, events, deadlines, comments/timer, documents, PDF,
transcripts, look, archive search and settings travel as fragments appended at the end of the prompt only when that
tool group is open, so the cached prefix survives. New base tool `open_tools` lets the model unlock a group itself
when no keyword opened it (no wasted retry). Env knobs `KB_CHAT_MAX_TAHU` and `KB_CHAT_MAX_TOOL_STARE` for
measurement only.
- **Assistant: confirming a simple card no longer calls the model.** After "Yes" on adding/changing/deleting steps,
ideas, sharing, preferences and similar actions in a free conversation, the app itself replies "Done." with the
"What next?" chips. The model still finishes after projects, rules, events, invitations, deadlines and in every
guide. `KB_CHAT_POTVRZENI=model` restores the old behaviour.
- **AI quota: 1 turn = one ordinary message.** The weekly quota is counted in turns where 1 turn is the average real
turn (≈ 17 600 input tokens, 65 % from cache, 284 output, priced as DeepSeek V4.1 at AKI); prices and the reference
turn live in one file, `pb_hooks/kredity-ceny.json`. Work of local models costs nothing; a discarded light-model
attempt, a failed turn and an app-side confirmation cost nothing. Credits are stored with each log row
(`ai_chat_log.kredity`, per-call breakdown in `volani`); older rows are recomputed from their token totals. The
Organization admin card and the quota messages say "turns" and "messages". Weekly usage shown to customers drops to
roughly half; the plan quotas are unchanged.
**Upgrade notes:** migrations add `ai_chat_log.skupiny`, `prepis_ms`, `kredity`, `volani` and recompute `kredity` for
existing rows. Nothing to configure; `KB_CHAT_POTVRZENI` is optional.
## v0.74-beta — 2026-10-05
**The assistant in the simplified phone view; "Open the full app" instead of "full version"**
- **AI assistant in the simplified view (`/lite`)**: an **Assistant** button in a new shared header
(logo · Assistant · Full app) on all three screens. The panel opens full-screen (also in a wide
window — the simplified view is a phone product), the down arrow closes it, and the same
conversation continues on the computer. The panel's chunk and its texts download only on the
first tap, so a cold `/lite` load costs nothing extra (`tests/lite-bundle.js`, new
`tests/ui-lite-asistent.js`). The server tells the model where the user is ("in the simplified
phone view… no map canvas"); the "Show in map" link opens the map in the full app without
changing the remembered view. The confirmation card for switching to the simplified view no
longer claims the assistant is missing there, and the way back it describes is now correct
(the view has no user menu).
- **"Full app" at the top too**: the way back to the full app used to be only below the list;
on a phone nobody found it. Now it is in the header on every screen, and the button below the
list stays.
- **Renamed "Switch to the full version" → "Open the full app"** (header: "Full app"). A user read
"full version" as buying a paid edition — the pricing page has a "Cloud Lite" plan. All texts
(hint, truncated overview, quick add without a project, the assistant's settings values and
cards, docs cs/en, README) drop the word "version". Internal names (`/lite`, `kb-mode`) are unchanged.
- **Lite size budget 515 → 560 kB** (`tests/lite-bundle.js`), rewritten from a +5 kB ratchet (raised
six times in five weeks) into a reserve with a stated purpose; the library ban and the
"lite < 70 % of the full app" rule remain the real protection. **`tests/scale-limits.js` measured
the slow-link ("3G") first visit through the service-worker cache** (cache-first for `/assets/`,
which `Network.setCacheDisabled` does not touch) — it now opens a fresh context with the service
worker disabled, so the numbers are a real first visit: the simplified view 4.2 s, the full app
11.5 s at 1.6 Mb/s without compression (the lite figure of 0.4 s was the cached one).
## v0.73-beta — 2026-10-04
**Share dialog with a member picker; fixes from the review panel after v0.72**
- **Share map: pick a colleague from the organization.** The dialog has a select of the instance's members above
the e-mail field — only people with an account who do not have the map yet, never yourself. Picking one fills
the e-mail; Invite and the rights work as before and typing an e-mail still works (also for addresses outside
the organization). Available from the editor and from the map cards on the home page (the home page used to open
the dialog without the member list). `ui-sdileni-vyber-clenu.js` covers both.
- **Assistant: language, the e-mail mode and turning e-mails or all notifications off go through a confirmation
card first** (`set_preference`, `set_notification`). An instruction hidden in a step description of a shared map
or in a document the assistant reads could otherwise change your language or silence deadline e-mails without a
click — the same reason the card already applied to turns with an attachment. Theme, readability, alignment
lock, names, turning a notification on and turning one type off in-app are still applied right away with
Revert; the card of a confirmed setting now carries Revert too.
- **Assistant: Revert sticks.** A reverted setting (language, lite mode, readability, notifications) was applied
again when another card from the same turn was confirmed — every settings card of the turn was re-applied on
each confirmation — so the browser ended up in English while the account was Czech. Each card is applied once
and never after Revert; the card shows "Reverted"; and when the revert does not reach the server the panel says
so instead of silently diverging. Settings cards now carry an id for this.
- **Assistant: confirming a card is one-shot even for two concurrent requests** (a second window, a retry after a
timeout). The pending card is removed atomically (`UPDATE … WHERE pending = <what was read>`) before the action
runs; the loser gets 404 and an invitation, a comment or a deletion is not performed twice. A batch containing a
browser-side card is rejected before any card of the batch runs (it used to fail in the middle).
- **Assistant: `move_node` creates the edge like the editor does** (`type: deletable`, so the delete button shows,
id `edge-…`) and sends `base_updated`, so a concurrent autosave of the open editor is detected (409) instead of
silently overwritten.
- **Assistant: the duplicate-title check is per branch, not per map.** The same step title under a different
parent ("Order material" under two orders) is legitimate; the parent, the steps above it and its whole subtree
still refuse a second node with the same title. The check now also covers `add_idea_to_map`, as the system
instructions promised.
- **Hosted instances: an AI agent webhook cannot point at a private or local address** (same rule as the AI
settings), both in the UI route and through the assistant's `save_ai_agent`.
- **Cloud chat proxy: health checks answer 401 without a token for unknown upstream names too** — the 401/404
difference allowed guessing upstream names anonymously. `cloud/tests/chat-proxy-tokens.sh` is now part of
`tests/run-all.sh` (cloud suites are listed by hand and this one was missing).
- Fixed: Organization admin → **Instance purpose** showed the raw keys `userAdmin.purposeTeam/Family/Solo` (the
dialog builds the key dynamically, so the unused-key clean-up had removed the texts); the texts are back.
- Tests: declining destructive cards (delete project, delete rule) leaves the data and tells the model; double
confirmation; `set_rule_enabled` and `list_rule_templates` (never called before); the per-branch duplicate rule
and the idea variant; the member picker on the home page; and a docker-free parity test of the CS/EN system
prompt (same keys, same `{placeholders}`, every tool name mentioned in the prompt exists).
- Housekeeping: one `normText` helper in `helpers.js` instead of three copies, dead org-structure reads removed,
lint warnings in the share dialog, the share dialog's member memo no longer recomputes on every render.
**Upgrade notes**
- No migration and no new variable. One new route, `POST /api/kb/chat/vratit` (the assistant panel records a
reverted settings card). The assistant's system prompt and two tool descriptions change, so the prompt-prefix
cache is rebuilt on the first turn after the upgrade.
## v0.72-beta — 2026-10-04
**The assistant handles every setting of the app**
- **The rest of the app's functions through the assistant** (`delete_node`, `archive_project`, `rename_project`,
`delete_project`, `set_map_public`, `update_rule`, `delete_rule`, `save_rule_template`, `delete_rule_template`,
`list_reminders`, `delete_reminder`, `request_deadline_change`, `decline_deadline_request`,
`mark_notifications_read`, `report_problem`): an inventory of every route and UI write against the assistant's
tools showed these gaps (the user's rule: everything except the security exclusions). Map/rule/reminder writes go
through the v1 API with a temporary key like the existing tools; project archive/rename/delete, the public link,
deadline requests, read-all and reports go through the app's own routes with the user's session token, so the
dialogs' permission rules apply (owner archives/deletes, editor renames, co-manager shares, assigner declines).
Every tool checks before the card (unknown rule, apex node, already archived, own project, no reminder…).
Second inventory pass added `add_comment`, `start_timer` / `stop_timer` / `get_timer`, `move_node` (with cycle
check), `update_idea`, `delete_document` and `revert_document`.
- **Review panel fixes before the release:** a direct settings tool now goes through a card in the whole conversation
that contains an image, voice or PDF transcript, not only in that turn (an instruction hidden in an attachment could
otherwise be executed one turn later); turning a notification channel off through the assistant keeps the e-mail
channel exactly as the app's own dialog does (the e-mail mode no longer leaks into the stored preference); the
keyword groups for settings were narrowed so "add a step to the project" or "order parts" no longer load 24 settings
schemas, and the deadline-request tools moved into their own keyword group; the share lookup uses one query instead
of one per map.
- **Assistant: a condition step goes under its step; no duplicate titles** (`feat/asistent-podkrok-podminka`): "before we
order we must measure" puts Measure under Order instead of beside it, and a step whose title already exists in the
map is not added again.
- **Cloud chat proxy: health check only with the tenant token** (`fix/chat-proxy-healthz-token`): the health endpoint
no longer reveals machine or vendor addresses, and the result is cached for 30 s.
- **Calendar events: update and delete through the assistant** (`update_event`, `delete_event`): move, rename, change
invitees, set/change/remove the reminder ("remind me an hour before" on an existing event), cancel; an invitee can
only leave the event. Found in the klik-test: "add a reminder" to an existing event ended in "I can't" and an offer
to delete and re-create it. The prompt now also says: a suggestion chip must be something the assistant can do, and
a click on it acts right away instead of asking again.
- **Project sharing through the assistant** (`get_map_sharing`, `share_map`, `unshare_map`, `set_team_access`):
"Add Petr to the Kitchen project", "Remove Karel's access", "Give the whole team edit access", "Who sees the
Kitchen project?". Same rules as the Share dialog — the owner or a co-manager (edit) shares by name, team access
is the owner's only, the org map is never shared this way; writes go through `/api/kb/share` with the user's
own session token, so every guard of the dialog applies. The prompt now tells the model to say right away when
it has no tool for a request instead of asking clarifying questions first (found in the klik-test: two
questions, then "I can't").
- **Personal settings through the assistant** (`set_preference`, `set_notification`, `get_settings`): language,
light/dark theme, simplified view, map readability, alignment lock, full and display name, notification
preferences per type or all at once, and the notification e-mail mode. Applied right away with a **Revert**
link on the card; the browser-only ones (theme, view, readability) are applied by the browser from the card
and reported to the server with each request so the assistant can read them. Switching to the simplified
view (no assistant there) always goes through a confirmation card that says how to come back; in a turn
with an image, PDF or voice note the language and notification changes go through a card as well
(instructions embedded in an attachment must not change them silently).
- **Organization settings through the assistant**, only for the roles that have them in the UI and always
through a **Yes, do it** card: `invite_member`, `update_member` (role, AI agents / org structure manager
flags, deputy — never the user's own role), `update_organization` (name, purpose), `set_ai_settings`
(own server only; a partial change is merged with the stored settings, the token is never sent),
`set_ai_credits`, `set_instance_skin` (performed by the browser, which holds the built-in skins),
`set_billing` and `order_membership` (hosted only), `save_ai_agent` / `delete_ai_agent` (no secret),
`get_org_structure`, `add_org_position`, `update_org_position`, `remove_org_position`. Members never
receive the administrator tools (their schemas are not even sent to the model); the role is checked again
when a card is confirmed.
- **Writes go through the same routes as the UI** with the user's own session token (`/api/kb/invite`,
`/purpose`, `/ai-settings`, `/billing`, `/member-deputy`, `/ai-agents/*`, `/org-structure/*`, the
`users` and `org_settings` records), so permissions, hooks (role lock, sanitised notification
preferences) and validation are shared — nothing is duplicated.
- **Deliberately not available**: password and e-mail change, API keys, the AI provider token, deleting an
account, resetting a colleague's password, the agent secret and the logo — the assistant points to the
menu item instead. An invitation on an instance without SMTP shows the temporary password **once** on
the card (with Copy); it is not stored in the conversation nor sent to the model.
- Docs: new section *App and organization settings* in the assistant guide (cs/en).
**Upgrade notes**
- The system prompt of the assistant changes once (a sentence about settings and where the excluded things
live); the prompt cache of running conversations is rebuilt on the next turn.
- New tools are offered by keywords (settings, language, notifications, invite, role, quota, agent,
structure…) and always on the Organization settings page; `KB_CHAT_TOOLS=all` offers them regardless.
---
## v0.71-beta — 2026-10-04
**Map unlocked on the phone, administrators' credit reserve, right page language**
- **Map on a phone: nodes can be dragged with a finger and connected**, also in the horizontal (phone)
layout, which used to be view-only. The lock is no longer on by default on small screens; the red lock
button stays as a manual choice. Positions are still stored in the vertical (canonical) layout: a move
made in the horizontal view is applied to the stored position with the axes swapped (down the row on the
phone = right along the row on a computer), so the order you see on the phone matches the computer. The
horizontal view itself is still laid out automatically when the map is opened.
- **AI credits: the administrators' share is a reserve, not a cap.** Administrators may use the whole
weekly quota of the organization; other members are capped at *quota − reserve*. A member is stopped at
the members' cap or when the whole quota is used; an administrator only when the whole quota is used.
`100 %` = other members get nothing, `0 %` = no reserve (one shared pool). The card in Organization admin
is labelled "Administrators' reserve (%)" and shows "reserved X, may use up to K · others at most Y";
`GET /api/kb/ai-kredity` returns `rezerva_admin`, the admin group's `kvota` is the whole quota.
- **`<html lang>` follows the real language** from the first byte (`cs` default, the stored or browser
choice via an inline script) instead of a hard-coded `en`. With `lang="en"` over a Czech UI some browsers
offered to "translate" the page into Czech and mangled the texts (reported by a trial user).
**Upgrade notes**
- No migration, no new environment variable. The saved "administrators' share" value is now read as a
reserve: administrators can use more than before (the whole quota), members at most *quota − share*.
`0 %` used to leave administrators without credits; it now means no reserve. Hosted trials keep
`KB_AI_KVOTA_TYDEN` as the hard cap.
- Server message key `err.aiKvotaAdmin` is replaced by `err.aiKvotaCelek` (organization quota used up);
`err.aiKvotaOstatni` now mentions the administrators' reserve.
## v0.70-beta — 2026-10-03
**Project numbers and an assistant that searches the archive**
- **Every project has a number** (`#12`): one sequence per instance, assigned by the server when a map is
created (a model hook — so it covers the app, the REST and v1 API, MCP, the assistant, templates, the
welcome maps and imports alike), never changed, never reused. It is a serial identifier for lookup, not
a label next to the title: shown in the project dashboard (and its PDF), in the tooltip of the map title,
in the Archive (the `#` badge there now means the project number — the position in a numbered series
stays in the title), by the assistant, and in `GET /v1/maps`, `GET /v1/maps/{id}`, the portfolio, MCP
`list_maps` / `get_map` and the data export (`project_number`). Project cards and the map header do not
show it. The organization map has no number. A number sent by a client is ignored; `PATCH` cannot change it.
- **Assistant: `search_projects`** — searches active **and archived** projects by project number, project
title and the titles, descriptions and owners of their steps (case- and accent-insensitive, word stems);
results are grouped by project with its number, state, open steps and deadlines (capped: 10 projects,
40 step lines). `get_map` accepts the project number (`"#12"`) and finds archived maps by title as well
(active ones take precedence). The map list in the system prompt carries the numbers; the prompt explains
when to search.
- **Archive page**: a search box filters the archived projects by title (accent-insensitive) or `#number`.
**Upgrade notes**
- The migration numbers existing maps by creation date (oldest = `#1`) with plain SQL — `updated` stays
untouched, open editors see no conflict. It also creates the `instance_settings` row with the counter when
missing. Numbers are not carried over by import; an imported map gets the next number of the instance.
- The system prompt of the assistant changes once (numbers in the map list, a sentence about the archive) —
the prompt cache of running conversations is rebuilt on the next turn.
## v0.69-beta — 2026-10-02
**Map toolbar that stays put**
- **Arrange, Sort and Readability are menus with a fixed label.** Arrange and Readability no longer cycle
on click (their label used to carry the current value, so the button changed width and the toolbar
shifted under the cursor): a click opens a list — the same pattern Sort already had — and the icon shows
the current value. Picking the style the map already has re-arranges it.
- **The style lock is a menu item** ("Lock style for every map") at the bottom of the Arrange menu;
press-and-hold is gone. The locked button still looks pressed in.
- **The frame button tidies the map by your settings**: it sorts by the chosen Sort key, lays the map out
in its Arrange style and zooms out to all of it (one Undo step; nothing is written when the map is
already tidy). On a map not yet arranged in this browser, on a read-only map, in the public view and in
kanban it only zooms out. The plain zoom-out stays in the canvas controls.
- **Nothing shifts the toolbar any more**: the "Saving… / Saved" indicator has a reserved slot in the wide
toolbar (it used to push every button ~77 px to the left after each save) and is a small icon outside the
flow in the icon toolbar (it used to push the direction switch ~65 px); toggled buttons (Assistant,
Bottlenecks incl. the reserved place for the amber counter, a locked Arrange) keep their width; and with
the wide toolbar (≥1850 px) an open assistant panel lies over the right end of the toolbar instead of
squeezing it — the logo and the back arrow no longer disappear. The icon toolbar (<1850 px) still fits
next to the panel as before.
- A click outside an open Arrange / Sort / Readability menu only closes it — it no longer also triggers
the node button under the cursor (e.g. "Add sub-goal").
- The style a map shows is no longer inherited from the previously opened map when you move between maps
inside the app.
- **Phone: both toolbar rows stay on screen.** The editor used `100vh`, which on a phone is the height
without the address bar, so the page could scroll by one row and the row with the logo, the back arrow
and the actions slid out of view. The editor now uses the visible height (`100dvh`) and the toolbar is
sticky.
- Fixed: the "What will you use killBottleneck for?" dialog (shown to the administrator of a new instance)
showed the raw keys `teamHint` / `familyHint` / `soloHint` under the choices since v0.46.1-beta — a
clean-up of "unused" translation keys had removed them because the dialog builds the key dynamically.
The texts are back and `ui-dotaznik-ucelu.js` now guards both the source and the rendered dialog.
- Fixed: a temporary throttle of the AI gateway (HTTP 429 from the per-minute limit or from the protection
in front of the gateway) is no longer reported as an exhausted quota. Trial instances used to say the
trial AI "is used up for this month" and paid ones a bare "The AI service rejected the request"; both now
get "The AI is busy right now. Please try again in a moment." (`429 ai_busy`, advisor route and voice-note
transcription; a voice note stays in the panel to retry). A really exhausted quota is reported as before.
No migrations, no new environment variables.
## v0.68-beta — 2026-10-02
**Meeting prep and Team meeting in the assistant**
- **Meeting prep** (AI helpers): the app asks at once (no model call) which meeting — the project the user
is in first, the nearest calendar events, recent projects, "With a person" (then the team's names as
chips); the model starts with the answer. New read tools `get_project_changes` (what moved in
a project over 7/14/30 days — the same logic as `/map-changes`, now shared via `mapChangeGroups`) and
`get_person_work` (a colleague's open work only in projects the user can see). Output: a draft with
Agenda / Open points / Stuck / What moved / To decide, then a calendar event on a card or an e-mail.
- **Team meeting** (administrators and managers only): new tool `get_team_work` over `buildPortfolio`
(team and shared projects only; private projects are neither listed nor counted). The app shows the team
overview (who has the most work, on fire, stuck) from the same data at once and asks *What shall we do
about it?*; after *Suggest handovers* the assistant proposes at most three handovers, writes them as `update_node` cards that may change only the assignee (and the deadline when
asked) and drafts a message for the team.
- Privacy layer of the team meeting: the role is checked at the start, on every turn and when a card is
confirmed (403 `err.teamMeetingManagerOnly`); only an allow-list of tools is offered and executed (also
through the tool-group safety net and on confirmation); no personal memory, project notes, map list or
page context (open map, selected node) reaches the model; reads and writes only in team/shared maps.
- `/map-changes` behaviour unchanged (the route now calls the shared helper).
- A draft saved without a title no longer takes a section heading as its name ("Program:" in Documents):
when the first line ends with a colon the title falls back to the kind, the project and the day
("Podklady na schůzku – Dílna 1. 10."); the meeting-prep prompt now asks for a title.
**After the meeting and Weekly review in the assistant, links from My day**
- **After the meeting** (AI helpers block): the app opens at once (no model call) with an invitation to
send a photo of the notes, a voice note or the text and who was there — naming today's meeting from the
calendar when one has already started; with the notes the assistant proposes tasks per project ("task → who, by
when", a new project, what is unclear), asks *Do it this way?*, then writes the tasks on cards and
drafts an e-mail to the participants "who – what – by when".
- **Weekly review**: a new read tool `get_week_review` — only the user's own work: finished in the last 7
days (from the change log), overdue, due or planned within 7 days, stuck, blocking others, assigned to
others and overdue. The app shows the overview Done / Stuck / Next week from the same data at once (no
model call) and asks what to handle next week (the tasks as options) and what to do with what is stuck
(Break it down · Write to the owner · Leave it); the answers go to the main model in the hybrid setup.
- Cards that assign a step to someone who cannot see the project yet say so up front: "(this gives … access
to the project)" — the assignment shares the project (work access), as the v1 API always did.
- **My day → assistant**: *Discuss in the morning briefing* below the morning encouragement and *Stuck?*
on an overdue task (breaks that task down with the assistant).
**AI helpers in the assistant — New project with AI replaces the old advisor dialog**
- An **AI helpers** block below Evening planning (empty conversation only): **Sort my notes** and
**New project with AI**. A grid on a computer, a collapsed row "AI helpers (2)" on a phone (the
choice to expand is remembered per account).
- **Sort my notes**: like Evening planning without "today" — paste a photo, record a voice note or
write ideas; the assistant sorts them and recommends a new project / into a project / keep in the
buffer, writes on cards after your Yes. The idea buffer panel has **Sort with AI** from two ideas
on (the assistant goes through the ideas already in the buffer).
- **New project with AI**: the **Or let AI draft the project…** link in the New project dialog now
opens the assistant with the goal you typed; the emoji, colour and client from the dialog are put
on the new map by the server (the model never sees them, the public v1 API is unchanged). The
assistant asks 3 questions, the app adds **How detailed should the plan be?** (Brief 5–7 points /
Detailed 3 areas × 2–3 steps / In-depth 3 levels, 18–25 steps — the rules of the former advisor),
checks the proposed tree against the chosen level before the card (at most twice, then lets it
through) and shows the whole tree on the card. Without a goal it asks for one or for material
(text, `.txt`/`.md`, PDF, photo, voice note). Started from the AI helpers block without a goal it opens
with a **form** like the former advisor (goal with clickable examples + level of detail), composed by the
app itself — instantly, no model call, no credits; the model then asks the follow-up questions. Whoever
types the goal into the message box (or sends a voice note) instead of filling the form gets the
level-of-detail question with the follow-up questions (asked at most twice per conversation), so the tree
is always checked against a chosen level. In the hybrid setup the follow-up questions come from the main
model, and an over-sized In-depth tree is returned with a concrete hint (how many sub-steps may stay).
- **Guided flows open instantly**: the opening step of the morning briefing, Evening planning and Sort my
notes (the invitation to send a photo, a voice note or ideas, with the choices on a card) is composed by
the app — at once, no model call, no credits. Tapping "I will write my ideas" (or "I will send a
photo…") gets the app's own "Go ahead, I am waiting for …" with a "Nothing to add" chip; the model starts
with your material or answer.
- **Delete ideas from the buffer through the assistant**: a new tool `delete_ideas` (exact titles, or the
whole buffer) — always a confirmation card that lists every idea to be deleted and says it cannot be
undone; only the user's own ideas, and exactly those the card showed. *Sort the buffer* now also suggests
a "Delete" section for obvious test or nonsense entries, and the assistant no longer offers to put ideas
"into the buffer" that are already there.
- Confirmation cards: new steps show their **description** on the card (e.g. a measurable goal written by
the assistant), and a rename reads *Rename "X" to "Y"* instead of "(title: …)".
- **Always something to click**: a reply that ends with no question, no card to confirm and no chips gets a
chip — the other options of the question just answered (e.g. "Nothing to add, go on" while the assistant
waits for your notes), otherwise "What next?".
- The confirmation card for `create_project`, `create_project_from_ideas` and `add_nodes` shows **the
whole tree** (indented, with owner, deadline and plan; the first 12 rows and *Show the whole
tree*). A plan date further than 7 days or an unknown owner is returned to the model **before**
the card instead of failing after your Yes.
- The chat file picker takes `.txt`/`.md` (the content goes into the input box); a message can be
8,000 characters (was 6,000).
- **Removed**: the old *Suggest with AI* / *Map from text* dialogs (`AiCreateDialog`,
`AdvisorDialog`, `FromTextDialog`) and their translations. The wand on a node, the project AI
summary and the morning encouragement stay. The docs page "Working with the AI assistant" now
points to the assistant page.
- Fixed: the project dashboard showed the raw key `editor:aiChat.connectionError` when the AI summary
failed.
**Voice notes in the assistant**
- A **microphone** next to the paperclip (on a phone in the bottom bar in place of *Send* while the
field is empty): record, **Send** — the voice note goes out right away, the app transcribes it and
the assistant sorts it like a photo of notes (changes on cards to confirm). The bubble shows
"🎤 Voice note 0:42" and the transcript; the recording itself is not stored. At most 5 minutes
(`KB_CHAT_HLAS_MAX_S`, recording stops and sends by itself), under 1 s is not sent; a finished
recording (a WhatsApp `.opus`, `.m4a`, `.mp3`, up to `KB_CHAT_MAX_AUDIO_MB`) can be sent with the
paperclip or dropped; a file that is too large is refused in the browser with a plain message. When
the voice note does not go out — transcription fails, the network drops, a proxy returns an error —
the recording is kept: *Try again* / *Download* / *Discard* (the server marks the one case where the
transcript is already in the conversation, `ulozeno`, so it is not sent twice). Recording runs only
while its bar is visible: closing the panel or switching to PDF stops it and keeps the note for you
to send — the microphone never runs blind.
- **Fix a transcript with the pencil**: the last voice note or photo transcript has a pencil in its bubble —
edit the text, *Send the correction*, and the assistant answers again from the corrected text
(`POST /api/kb/chat/oprav`; unconfirmed proposals of that turn are dropped, one more AI turn). Offered
only while nothing has come out of that turn yet (a confirmed card, a saved document, memory, an idea);
the conversation DTO carries `lze_opravit`.
- **Enter sends the answers** from the own-answer field of a question card (with several questions it
first jumps to the next unanswered one).
- The microphone shows only where transcription works (`/api/kb/config` → `chat_voice`,
`chat_voice_max_s`, `chat_voice_max_mb`). Browsers allow the microphone only on https/localhost —
on a plain-http address tapping it explains why; attaching a finished recording still works.
Morning briefing and evening planning invite a voice note too.
- **One transcription path** (`pb_hooks/prepis.js`) for voice notes and the transcription API mode
(`/api/kb/advisor`, `mode: transcribe` — the old *Upload audio* dialog is gone, see below):
`KB_TRANSCRIBE_PROVIDER=openai` + `KB_TRANSCRIBE_URL`/`_MODEL`/`_TOKEN` (OpenAI shape — speaches,
whisper.cpp, OpenAI…; `none` = off), otherwise the general AI as before. `KB_AI_TRANSCRIBE_URL`
ending in `/audio/transcriptions` now really gets the OpenAI shape (the docs promised it); the
service key goes there only on the same origin. The recording type is recognised from its
content (a WhatsApp `.opus` goes as `.ogg`). A voice turn uses 2 turns of the hourly cap
(`KB_AI_HLAS_VAHA`), recordings up to `KB_CHAT_MAX_AUDIO_MB` (3). The length is logged
(`ai_chat_log.audio_ms`); credits are still charged only for the assistant's reply. The
transcription API mode has its hourly cap with every provider (`KB_AI_MAX_TRANSCRIBE_PER_HOUR`),
and a failing transcriber no longer returns its address in the error text.
- Android app: `RECORD_AUDIO` + `MODIFY_AUDIO_SETTINGS` permissions.
**Fixes from the pre-release review**
- **Deleting ideas only ever deletes what the card showed**: a stale id (the idea was removed elsewhere,
or moved by another card of the same turn) no longer falls back to a title search that could hit a
different idea with a very short title. Two ideas with the same title are both listed and both
deleted; the result tells the assistant how many ideas remain.
- A proposal that would not fit into the confirmation queue (a big tree with long descriptions) is
returned to the model to shorten or split instead of losing the whole turn.
- Cards name the **real project** even when the model passed only part of its title; *Create a project
from ideas* says who gets access through an assignment, like the other cards.
- Team meeting: `null` in optional fields (GPT models send it) is no longer taken for a forbidden
change; "Who has the most work" is ordered by open work, not by delay; an external contact shows by
name (or as "external contact"), never as its internal pseudo-address.
- A PDF sent as the answer to a question (the New project form invites one) reaches the model whole,
not cut at 8,000 characters; correcting a photo transcript no longer cuts the PDF text attached to
the same message, and the correction is saved in a single write (nothing is lost if the process
dies mid-turn).
- Weekly review: an idea from the buffer with a deadline is marked as such and is not offered among
the tasks to plan. After a voice note, "Create a new project" is added to the first question only
where sorting happens (not in Meeting prep, Weekly review or the Team meeting).
- Hybrid setup: the first model turn of Meeting prep goes to the main model also after "With a
person → name". Fallback chips no longer offer the other options of a question that has just
been carried out ("No, thanks" after the notes were saved).
- Limits: the hourly weight follows the real mode (an unknown `mode` no longer made an image or
voice turn count as 1), a nonsensical voice-note length from the client cannot drop the usage log
row, `POST /api/kb/chat/oprav` has a body limit.
- Panel: question cards and an open transcript editor no longer carry their state over to another
conversation picked from the history; a reply that arrives after you switched conversations is not
drawn into the wrong one; the "Looked into: …" line names the new read tools.
- **An error in a running conversation is shown again**: when sending or confirming failed, the
message explaining why was cleared at once by the reload of the conversation (it only stayed in a
brand-new conversation) — the message just vanished without a word. In the app since the assistant
panel exists.
- A transcript longer than a message can hold (an attached recording of tens of minutes) ends with a
note that the rest did not fit, instead of being cut silently.
**The assistant on the real OpenAI API**
- **Fixed: the conversation stopped with HTTP 400 on api.openai.com** whenever the assistant asked a
question and offered next-step chips in the same reply (the morning briefing and evening planning
do exactly that), when you typed instead of confirming a card, or confirmed one of two cards. The
history sent to the model now answers every tool call before anything else, as OpenAI requires
(Ollama, llama.cpp and most gateways tolerated the old order, so it went unnoticed).
- The last round of the tool loop and the "answer in text" retry no longer send an empty `tools`
list (OpenAI rejects it).
- **Reasoning models** (o-series, `gpt-5*` incl. GPT-5.6 Luna): the request goes out in the shape they
accept on the first try (no double call), with a reserve for reasoning (`KB_AI_REASONING_BUDGET`,
default 4000 tokens) and `reasoning_effort` (`KB_AI_REASONING_EFFORT`, default `low`; `auto` = do not
send; `KB_*_OPENAI_EXTRA` wins). A model whose name does not reveal it is recognised after its first
refusal of `max_tokens`/temperature or by `reasoning_tokens` in the reply, and remembered. Applies to
the assistant, the old advisor functions, summaries and image reading.
- GPT models get all assistant tools at once (`KB_CHAT_TOOLS=auto`, default) — they do not call a tool
they were not offered, so keyword-gated features silently disappeared. Qwen, DeepSeek and gpt-oss
keep the tool groups. `null` in optional tool arguments is ignored instead of failing validation;
Markdown in replies (`**bold**`, `#` headings, `*` bullets, tables) is converted to the panel's
plain-text style; a model refusal is shown as the reply.
- **Images only where they can be read**: `/api/kb/config` reports `chat_image` when a vision model is
configured. Without it the paperclip takes PDFs only, a pasted image gets a local explanation, and
the morning briefing / evening planning invite you to write your ideas instead of pasting a photo
(with a vision model the wording is unchanged).
- **Images from Administration**: with an Ollama or OpenAI provider set in Administration → AI, tick
*The assistant reads images* (optionally a separate image model) and save — the app immediately
tries a built-in image with known text and enables images only if the model actually reads it
(*Test an image* repeats the check; the test always uses the saved settings). Until now images
needed the `KB_VISION_*` environment variables, so an OpenAI key entered in the app gave no images.
### Upgrade notes
- New migration adds `vision_enabled`, `vision_model`, `vision_ok` to `ai_settings` (defaults off)
and `audio_ms` to `ai_chat_log`.
- The `/api/kb/chat` body limit grew from 2 MB to about 4.25 MB (a 3 MB voice note in base64). **If a
reverse proxy sits in front of the app, raise its request body limit accordingly** (nginx:
`client_max_body_size 5m;` — its default of 1 MB rejects voice notes and larger photos with 413).
- **Drafting a project with AI now needs the assistant** (provider `ollama` or `openai`, or
`KB_CHAT_*`). Instances whose AI runs only through the remote `api` service or a `custom` endpoint
lose the *Suggest with AI* / *Map from text* dialogs without a replacement; the wand on a node, the
project AI summary and the morning encouragement keep working there.
- Rolling back to v0.67-beta works over the same data (tested): the new fields are ignored and
conversations started with the new helpers continue as ordinary conversations.
- `KB_CHAT_TOOLS` default is now `auto` (see above); set `groups` to keep the old behaviour for GPT models.
---
## v0.67-beta — 2026-10-01
**Assistant documents — notes, e-mails and summaries next to the chat**
- Everything the assistant writes with `draft_text` (e-mail, meeting / call points, note, summary)
is now saved as a private **document** (new collection `ai_documents`, only the owner sees it).
A **Documents** panel opens to the left of the assistant, over the map (the page is not pushed
aside): list with filter All · Notes · E-mails · Summaries and search, detail with Edit, Copy,
Open in e-mail (`mailto:` — nothing is sent by the app), Download as text, Restore previous
version and Delete. Only the assistant creates documents (ideas and tasks stay in the idea
buffer, the prompt draws that line). On a computer the panel opens by itself on a document
the assistant just wrote or changed; on a phone it covers the screen with Back to chat.
- New assistant tools `list_documents`, `get_document` (read) and `update_document` (rewrites a
document, keeps the previous version; in turns with an image/PDF it needs a confirmation card).
`draft_text` gains kinds `note` and `summary` and fields `subject` / `to`; a "Subject:" first line
of an e-mail is split off automatically.
- Morning briefing and Evening planning end with "Save the notes to documents?" (a summary).
- A document that belongs to a project shows the project in the list and an **Open project** link
to its map (only while you can see that map).
- **Memory moved into Documents**: the brain button in the assistant header is gone; "What the
assistant remembers" is the pinned first item of the Documents panel (and "Open memory" under the
memory card in the chat).
- **Drafts are no longer appended to project notes** — the "Save to project" button and the route
`POST /api/kb/chat/koncept-uloz` are removed; `draft_text` with `map` only links the document to
the project. Project notes (sent to the model in every turn on that map) keep short facts only.
Drafts already appended earlier stay in the notes untouched.
- Routes `GET /api/kb/chat/dokumenty`, `GET /api/kb/chat/dokument/{id}`, `POST /api/kb/chat/dokument`
(edit only), `POST …/dokument/vratit`, `POST …/dokument/smazat`. Limits: 200 documents per
user, 20,000 characters each. Included in "Download all my data" (`documents`). Not exposed in the
v1 API / MCP.
**Upgrade notes**: one migration (`1790801929_ai_documents.js`) adds the collection `ai_documents`;
nothing else changes in existing data. Drafts already appended to project notes stay there. Rolling
back to v0.66-beta is just the previous image — the new collection then sits unused (never run the
down-migration by hand, it deletes the documents).
---
## v0.66-beta — 2026-09-30
**Morning briefing asks for your notes + Evening planning**
- **Morning briefing** now opens by inviting you to paste a photo of your notes (Ctrl+V, drag and
drop, paperclip / camera on the phone) and to write down everything on your mind; the items are
sorted into the idea buffer or projects (cards to confirm) before the day's recommendations.
"Nothing to add, go on" skips the step. The offer on first open of the day is unchanged.
- **Evening planning** (`mode: nocni`): its own box in an empty conversation, below the
morning-briefing offer — always there, no day logic, no reminder, nothing to dismiss. The
assistant invites you to empty your head (photo + ideas), saves nothing right away, sorts the
items and recommends what to make of them (new project / into a project / buffer for later),
asks "Do it this way?" and only then creates the project and writes (cards). It does not touch
today's or tomorrow's tasks — the morning briefing does. History shows it with a ☾ icon.
`POST /api/kb/chat` accepts `mode: "nocni"` next to `porada` and `rozbor`.
- **Image transcript rules**: a heading above a list is a project title, not an item; app controls
(e.g. "Add item") are skipped; an empty checkbox is not "done"; pasted lists are sorted with a
recommendation before anything is saved.
## v0.65-beta — 2026-09-29
**Phone fixes from real use + "Do all" in the assistant**
- **Map header on the phone**: the row of eleven buttons could not wrap, overflowed the screen and
let the whole page (map included) scroll sideways. Below 640 px it is now two full rows — back,
add goal, bottlenecks, assistant, bell and menu on top; map tools (direction, align, readability,
fit) below. Desktop layout is unchanged.
- **Assistant on the phone**: a button at the top of the map (the same as on Home, highlighted like
*Add*); the panel shrinks by tapping the robot icon on the left of its header as well as the arrow
on the right. Toasts on the phone used to sit *over* the panel header for a few seconds after every
executed action, swallowing taps — with the panel open they now appear below the header.
- **Assistant "Do all"**: when several actions are proposed at once (e.g. a date and an assignee on
five steps), a *Do all (N)* button sits above the first pending card; individual Yes/No stay.
`POST /api/kb/chat/potvrdit` accepts `action_ids: [...]` next to `action_id` — actions run in card
order and the model replies once after all of them; an unknown id rejects the whole batch (404).
Browser-side actions (PDF fixes) keep their own selection and are not batched.
- **Left rail tabs on Home and Tasks** (idea inbox, timer, report a problem) were pinned at fixed
offsets computed for a one-line header without the trial banner; on the phone the inbox tab sat on
the *New* button. They now measure the real header height (banner included).
**Fixes from the second code review (wave A) — no new features**
- **Timeline**: tasks attached to a goal (and their subtasks) were never drawn on the timeline —
only goals and unattached tasks were; subtasks of unattached tasks were missing as well. Both are
shown now (the timeline compared the wrong node id since v0.57).
- **Assistant**: pressing Enter while the assistant was still thinking silently discarded the message
you were typing and any attached image/PDF — the draft now stays in the box. Morning-briefing
preferences are stored per account (a second person on the same browser got no briefing).
A member no longer sees the internal address of the AI gateway in a "did not respond" error
(admins still get the full text). `KB_AI_OPENAI_EXTRA` (e.g. `reasoning_effort`) is now passed to
the Advisor, the assistant and summaries (before, only summaries with env config got it).
- **Admin**: an admin creating a user through the REST API got a generic 400 (a JS scoping bug since
v0.47); deputy validation errors are readable again.
- **Notifications**: *organization notice*, *password reset* and *digest overflow* rows showed a raw
translation key instead of the type name.
- **MCP over HTTP**: `serverInfo.version` reported `0.1.0`; it is now the instance version (same as
`/api/kb/config`).
- **Operations**: cleanup crons no longer swallow errors silently (PocketBase logs them);
`docker-compose.yml` now passes `KB_PURPOSE_ASK`, `KB_AI_OPENAI_EXTRA`, `KB_CHAT_TOOLS` and
`KB_VISION_ZALOHA_NUM_CTX` (they were documented/read but not forwarded); new
`cloud/kontrola-env.sh` keeps env × compose × docs in sync; the public-export guard no longer misses
internal machine paths; build warnings are no longer suppressed (`logLevel`).
- **The "With AI" button is gone** (header split button, its menu item and the empty-state button on
Projects): creating a project with AI is the assistant's job now (side panel, `create_project`).
The Advisor dialog itself stays reachable through the **"…or let AI draft the project"** link at the
bottom of the New project dialog (Suggest with AI, Map from text, dictation). Three UI suites that
drove the removed button (`ui-ai-dialog`, `ui-ai-outage`, `ui-ai-timeout`) are retired; `ui-ai-mapa`
now goes through the link. Goal breakdown (wand) is unchanged.
- Dead code removed (`OperationsCard`, `advanceDate`), root README rewritten, `sync/REVIEW.txt` gone.
**Upgrade notes**
- If you keep your own `docker-compose.yml`/override, forward `KB_PURPOSE_ASK`, `KB_AI_OPENAI_EXTRA`,
`KB_CHAT_TOOLS` and `KB_VISION_ZALOHA_NUM_CTX` (they were read by the server but not passed through).
- `KB_AI_OPENAI_EXTRA` now also applies to the Advisor and the assistant (unless `KB_CHAT_OPENAI_EXTRA`
is set), not only to summaries — check it if you set it for summaries only.
- The "With AI" project button is gone; its replacement is the assistant, which is available only with
`KB_CHAT_*` configured (ollama/openai). An instance with the Advisor only keeps the wand (goal breakdown)
and the "…or let AI draft the project" link in the New project dialog.
- After logging out (or when another account signs in), the assistant's last-chat key is cleared; panel
open state, width and the model choice are kept.
- Three UI suites were retired (`ui-ai-dialog`, `ui-ai-outage`, `ui-ai-timeout`); forks running
`tests/run-all.sh` will see three suites fewer.
## v0.63-beta — 2026-09-20
**Events in the calendar, timed deadline reminders, and the assistant that sets both**
**Events in the calendar**
- The calendar's **+** now asks what to add: a **Task in a project** (a goal with that deadline,
as before) or an **Event**. An event is a meeting, the dentist, a video call — something with
a time that belongs to no project. It has a title, a day, a time (empty = all day), a note,
**invited colleagues** (instance members only; each sees it in their own calendar and gets an
*event invited* notification; an invitee can leave the event from its detail) and a **reminder** (at the start / 15 / 30 / 60 min / a day
before; an all-day event reminds in the morning at `KB_DEADLINE_HOUR`).
- A deliberate decision: **an event is the first item outside projects** (like the idea stash)
and **it is not a task** — no assignee, no status, never written into a map, changes nothing
in any project. It lives in the calendar only, under the always-present *Events* filter row, with the time
before the title on its chip. Dragging it to another day asks *"Move the event from X to Y?"*
and moves it (the creator only; invitees read).
- Today's events also appear in **My day** as a *Today in calendar* row ("14:00 Dentist"; plain
text in the lite view). Clicking opens the calendar with the event's detail.
**Timed deadline reminders**
- The goal detail of a goal with a deadline has **Remind me of the deadline** under the Deadline
field: on the deadline day / the day before (default, 16:00) / 2 days / a week before, plus the
hour. The reminder is **private** (one per person and node), **never changes the deadline**,
and **follows the deadline** when it moves; a done or deleted node drops it. A node with a
reminder shows a bell on its calendar chip.
**Notifications**
- Two new types: **Timed reminder** (`reminder`) and **Event invitation** (`event_invited`).
Clicking a reminder opens the calendar with the event or the map with the node.
- A minute-by-minute `reminders` cron delivers timed reminders to the bell and **by e-mail**.
For **your own events and node reminders** e-mail is **on by default** (a meeting reminder that
never reaches a closed app is useless) and it is sent **immediately even in daily-digest mode**;
the *no e-mails* mode still applies. An **invitee** gets the reminder in the bell; by e-mail only
after ticking *Timed reminder → e-mail* in notification settings (an invitation you did not ask
for must not be a way to send you e-mail). An event created or moved into the past is marked
as reminded — nothing fires retroactively. Reminder times are the **instance time zone** (`TZ`); the event dialog warns
when it differs from the browser's. After an outage, reminders older than
`KB_REMINDER_CATCHUP_H` hours (default 48) are logged, not sent.
**AI assistant**
- The assistant creates an event from one sentence ("dentist tomorrow at 2, remind me half an
hour before, invite Jane" → a confirmation card; unknown invitees are refused before the
card), lists your events ("what meetings do I have this week?") and sets a timed deadline
reminder ("remind me about the Novak quote the day before the deadline at 9" → a card; the
node must have a deadline — otherwise it first offers to set one). New chat tools
`create_event`, `list_events`, `create_reminder`; untimed alerts still go through a
`deadline_approaching` rule.
**API and MCP**
- REST v1: `GET/POST /api/kb/v1/events`, `POST /api/kb/v1/events/{id}` (owner only; an invitee
gets 403, an invisible event 404), `POST /api/kb/v1/events/{id}/delete`, `POST /api/kb/v1/events/{id}/leave`
(an invitee removes themselves);
`GET/POST /api/kb/v1/maps/{id}/nodes/{nodeId}/reminders` (upsert `{offset_days 0–30, time
"HH:MM"}`; a node without a deadline or a time in the past → 400) and
`POST …/reminders/{rid}/delete`. No `base_updated` on any of them — nothing in a map changes.
- MCP: 20 tools — new `create_event`, `list_events`, `create_reminder` in both the HTTP `/mcp`
server and the npm stdio package.
- **Hint change:** an unknown `reminder` / `remind_at` / `time` / `hour` field on a node now points
to the reminders endpoint / `create_reminder` (it used to say "create a `deadline_approaching`
rule" — that stays the advice only for untimed alerts or a whole map); `event` / `meeting`
point to `/v1/events` / `create_event`.
**Upgrade notes**: two migrations (`1789820000_udalosti_pripominky.js`, `1789889632_node_reminders_map_index.js`)
add the collections `events` and `node_reminders` (server-written only), an index, and the notification
types `reminder` and `event_invited`. Rolling the migrations back **deletes** both collections (events and
reminders are lost). The *Export all data* file now includes events and node reminders, but
*Import all data* does not restore them yet (it restores maps and the idea stash, as before).
Limits: 2,000 events per user, 60 saves per minute. **`TZ` now matters for reminders**: a self-hosted instance defaults to UTC, so a
reminder set for 14:00 fires at 14:00 UTC — set `TZ=Europe/Prague` (or your zone) before people
start relying on reminders. New optional variable `KB_REMINDER_CATCHUP_H` (default 48). API
integrations that sent `reminder` to a node endpoint still get a 400, but the hint now names
the reminders endpoints instead of a rule. Nothing else changes; existing rules and deadline
digests work as before.
---
## v0.62-beta — 2026-09-19
**PDF in the assistant: merge, split, extract, remove — and correct text with the assistant**
- **PDF in the assistant** (new **PDF** tab in the panel): merge several PDFs in a chosen order,
split by page ranges, extract or remove pages — all **in the browser**, the file never leaves the
user's computer and nothing is sent to the server or the model. **Correct text with the
assistant**: the page text (only the text) goes to the assistant, it proposes replacements
(price, name, sentence) on a confirmation card, the browser then edits the PDF as an overlay
(built-in Liberation font, original text stays underneath), shows a page preview and offers the
corrected file for download; the assistant reports honestly what was and was not replaced.
Limits: 60 pages / 40 000 characters of text per attachment, 30 MB per file; scans without a
text layer and encrypted PDFs are refused. A turn with a PDF weighs 2 in the hourly limit
(`KB_AI_PDF_VAHA`). New dependencies: `pdf-lib` (MIT), `pdfjs-dist` (Apache-2.0), Liberation
fonts (OFL) — loaded lazily, not part of the lite bundle.
- **Corrections add up**: every further correction card applies all earlier confirmed corrections
of the same file plus the new ones (the list is kept with the conversation, so it survives a page
reload — you only pick the file again). Corrections on the same line are drawn as one; unchecked
replacements are reported to the assistant as skipped. The page text of a PDF is treated as data:
memory or idea-buffer changes from such a turn need a confirmation card, like with images.
- The PDF tab keeps its file list when you switch to the chat and back; a locked PDF (password or
editing restriction) is refused right when attached, with a plain message.
**Upgrade notes**
- No migration. New optional variable `KB_AI_PDF_VAHA` (default 2). The PDF libraries are served
from the instance itself (no CDN) and load only when the PDF tab or a PDF attachment is used.
---
## v0.61-beta — 2026-09-17
**AI assistant on the side (with images and deadlines); share invitations for addresses without an account**
This is the first public release of the AI assistant. It was built and tested in two private
builds (`v0.61-beta-ai1`, `v0.61-beta-ai2`); everything from them is listed below.
**Sharing: invitation e-mail for an address without an account**
- Sharing a project with an address that has no account yet now sends an **invitation e-mail**
(subject names the person sharing, Reply-To goes to them, button leads to registration with the
address pre-filled). Where self-registration is not possible (registration key, hosted instance,
seat cap) the e-mail advises asking for an account. Previously nothing was sent at all.
- Anti-spam: one invitation per address and project, a daily cap per sender
(`KB_SHARE_INVITE_DAILY_CAP`, default 20) and per instance (`KB_SHARE_INVITE_INSTANCE_CAP`, default 50).
- The share dialog shows the invitation state, marks members "no account yet", offers a copyable
registration link when e-mail is not configured, and lets administrators/managers invite the
person into the organization directly.
**AI assistant: images, deadlines and safer rules**
- **Assignee for steps with a deadline**: before the assistant writes new steps with a deadline
that you would handle (or that have no assignee), it asks *"Do you want to be the assignee of the
steps with a deadline? Then you will see them in My day."* — enforced by the server, not only
by the prompt. "No" leaves them unassigned; steps for someone else need no question. The card
for adding nodes to an existing project now lists the assignees too.
- The assistant panel header has a labelled **+ New** button right next to the conversation title.
- Tasks page: the views are now ordered **Table, Calendar, Timeline, Kanban**.
- **Images in the assistant**: paste a screenshot (Ctrl+V), drop it on the input or use the
image button — e.g. a photo of handwritten notes, a phone task list or an e-mail. A vision
model transcribes it; the assistant then offers to put the items into the idea buffer, into
an existing project, or to create a new project (that option is always offered). The original
image is **not stored**: the conversation keeps the transcript and a small thumbnail (the last
3 per conversation). Several items go into the idea buffer on **one** confirmation card that
lists all of them. Configuration: `KB_VISION_*` (see `.env.example`); without it images are
refused. Text read from an image is treated as data: memory changes from such a turn need a card.
- **Deadlines through the assistant**: the assistant may now **set, change or remove a deadline**
(when a date agreed with someone follows from the conversation or a pasted e-mail) — always on
a confirmation card showing the old and new date. This replaces the v0.61-beta-ai1 note that
deadlines can never be changed by the assistant. Existing permission checks still apply.
- **Reminders that would never fire are not created**: a "deadline approaching" rule on a node
without a deadline, with a reminder day already in the past, or notifying the owner of a node
that has none, is rejected before the card with an explanation. Rule shape is validated before
the card too, so you no longer confirm a rule that then fails. After creating a rule the
assistant states exactly when the notification arrives.
- **Assistant behaviour**: no duplicated final answers, no running commentary between reading
steps, a project from pasted items is created directly (no detour through the idea buffer),
an "Open project" button under the answer, "what next" chips stay clickable.
- **Limits**: the chat route rejects bodies over ~2 MB before reading them; an image turn uses 3
turns of the hourly limit (`KB_AI_IMG_VAHA`), counted atomically and only for valid images.
**AI chat on the side: an assistant that sees your projects, asks short questions and proposes changes you confirm**
- **Assistant panel** (right edge, full height, collapsible to a tab, resizable 320–640 px): chat
with a model that can read your maps, My day, the idea buffer, rules and the organization
overview. Reading happens immediately; **every change (put an idea into a project, create a
project from several ideas, add or update nodes, create or toggle a rule) is shown as a card
and executed only after you press Yes**. Small personal actions (a single idea into the buffer,
a memory note, switching the look, a text draft) happen without a card — unless the turn
contained an image.
- **Short questions with prepared answers**: when the request is ambiguous the assistant asks
1–3 questions with 2–4 options each (plus a free-text answer).
- **Look switching**: "switch my look to sepia" changes the skin (with Revert); light/dark too.
- **Memory**: the assistant keeps markdown notes about you (preferences, context) — visible and
editable in the panel; you can clear it any time.
- **Configuration**: `KB_CHAT_PROVIDER/URL/MODEL/TOKEN` (own model for the chat), otherwise
`KB_SUMMARY_*`, otherwise the general AI settings. Needs a model with tool calling (ollama:
gemma4, qwen3.x, gpt-oss; OpenAI: gpt-4o-mini…). Usage is logged per model
(`GET /api/kb/chat/spotreba`).
- Writes go through the product's own v1 API with a temporary key of the user ("a key acts as
its owner"), so validation, rights and notifications are the same as for MCP and the app.
- Not in the phone (lite) layout in this version.
- The idea buffer panel (Home, Tasks, map editor) reloads itself as soon as the assistant adds an
idea or moves one into a project — previously the new idea appeared only after a page reload,
so "saved to the buffer" looked like a false claim.
- In the map editor the assistant also knows which node you have **selected**: "break this
step down" means the selected node. The client sends only the node id; the server looks the
title up in the map and appends it at the very end of the system prompt, so clicking through
nodes does not invalidate the cached prompt prefix.
- **Map shows "when I want to work on it"**: a node planned from My day (or by the assistant)
carries a small *Plan 14 Sep* badge next to the deadline badge, so the plan is visible in
the map too. Done nodes do not show it. The deadline badge is unchanged (a plan never moves
the deadline).
- When you state a goal or a problem ("I'd like to work with fewer interruptions"), the
assistant no longer only schedules: among the options it always offers **"Advise me how to do
it"** and, when chosen, gives 3–5 concrete steps tied to your map and offers to write them in
as sub-steps.
- Notifications (toasts) no longer cover the chat input while the assistant panel is open on
desktop — they shift left of the panel.
- **Cheaper prompts**: the system message now holds only what does not change between turns
(rules → date → mode → project notes → memory → map titles). Where you are and the selected
node travel as a short bracket in front of your own message and are stored with it, so the
history never changes retroactively; steps already offered are visible to the model in its
earlier `suggest_next` calls instead of a growing list. Measured on the real model: a state
change (task done, new suggestion) used to drop the cached prompt prefix to ~4k tokens and
recompute the whole history; now the prefix survives. The log records cached prompt tokens
(`ai_chat_log.tokens_cached`, from `usage.prompt_tokens_details.cached_tokens` or llama-server
`timings.cache_n`); the usage endpoint and the AI credits section count cached input at the
cache price. The map list no longer carries open-node counts (the model reads them with tools).
- **New project from scratch**: "I'd like a new map for a hot-dog stand" now makes the assistant
propose the title, the goal and 5–8 first steps itself and offer them on one confirmation card
(tool `create_project`). You are always the owner — it no longer asks for an owner e-mail —
and right after creation it offers fitting preparations (financial overview, suppliers…).
- The confirmation card for "add nodes" now shows the total count when the assistant proposes
more than four nodes (including nested ones), so you know what you are approving.
**AI credits in Organization admin**
- New section **AI credits** (administrators): this week's assistant usage in credits for the
whole organization and per person, split into *administrators* and *other members*, plus the
previous weeks. 1 credit ≈ one morning briefing (2 410 input + 454 output tokens, about
CZK 0.08 at the reference price); counted from the assistant chat log.
- **Weekly quota**: credits per week (0 = no limit) and the administrators' share in % (default
30 %); the rest is shared by the other members. Runs Monday to Sunday (UTC), no carry-over.
When a group's share is used up, the assistant answers with a clear message instead of a
reply. Hosting can cap the quota with `KB_AI_KVOTA_TYDEN`; the value saved in the app can
only lower that cap, never raise it. The check runs before a turn and usage is logged after
it, so the last turn may overshoot slightly (soft cap). Endpoints `GET /api/kb/ai-kredity`, `POST /api/kb/ai-kredity/nastaveni`.
- Not counted yet: the older AI features inside the map (they do not log tokens).
**Upgrade notes**: new collections `ai_chats`, `ai_memory`, `ai_chat_log` (server-written only);
new JSON field `instance_settings.ai_kredity`.
The panel appears only when the chat model is configured; nothing changes otherwise.
Images additionally need `KB_VISION_*` (see `.env.example`); sharing invitations need a configured
e-mail gateway. No other configuration change.
---
## v0.60.1-beta — 2026-09-15
**Wider pages on a large monitor (matching the calendar), top bar aligned with content**
- On a wide monitor the Projects, Templates, Tasks, Organization and Archive pages are wider —
matching the calendar — and the top bar lines up with them. Nothing changes on smaller windows
or on a phone.
Upgrade notes: no migration, no configuration change.
---
## v0.60-beta — 2026-09-11
**New calendar on Tasks: Month / Week / Day / Agenda, drag a deadline with confirmation, phone layout; goal icons in My day**
- **Calendar (Tasks page)**: rebuilt from scratch. Four views — Month, Week, Day and Agenda —
a side panel with a mini calendar, project and status filters, search, a "+" on every day to
create a goal with that deadline, and a day detail listing everything due. `/tasks?view=calendar`
opens it directly. Overdue items carry a red dot, done ones are struck through; colours follow
the skin, projects use their map colour.
- **Drag a deadline**: drag a chip to another day (Month) or column (Week). The app asks
*"Change the deadline from X to Y?"*, writes the new date only after confirmation — with a
compare-and-set guard so a colleague's concurrent change is never overwritten — and the toast
offers **Undo**. Done items don't drag. Someone who may not change the deadline (not the
assigner or map owner) gets a *deadline request* dialog instead; the request goes to the
assigner and can be withdrawn from the toast. Permissions are enforced on the server as before;
the client only avoids letting you drag into an error.
- **Phone (< 640 px)**: the Month fits the screen with status dots, tapping a day opens its detail;
the Week becomes a vertical list of days with a header per day; the day detail fits the display.
The Tasks view switcher now wraps on narrow screens (it used to push the page wider than the
screen, so fixed dialogs were centred off-screen).
- **My day**: goals show their icon (emoji) instead of the generic target when one is set.
- Keyboard: `t` today, `m`/`w`/`d`/`a` views — ignored while typing, in dialogs and during a drag;
Escape closes the day detail. Collapsing the side panel keeps the grid (it used to hide the whole
calendar).
- New suites: `ui-kalendar` (57 checks incl. phone), `calendar-dates` (129, three time zones, DST,
end-of-month clamping), `kalendar-data` (30, the move matrix and permissions); `my-day-api` and
`ui-myday-node` extended for icons. Lite bundle 504 → 508 kB (cap 510).
Upgrade notes: no migration, no configuration change. The old `calendar.*` texts moved to a lazily
loaded namespace; instances with a custom translation overlay should re-check the Tasks page.
## v0.59-beta — 2026-09-06
**Arrange by deadline / plan / assignee / status; bottleneck counter "1 + 3"; Rose skin**
- **Sort (map editor)**: a new dropdown next to Arrange (align) sorts the *siblings* under every
parent by deadline, plan (`plannedOn`), assignee or status. Deadline and plan use the
earliest date in the whole branch (a category with an urgent sub-goal moves first),
empty values go last, assignees sort by display name, status runs in progress → not
started → done. Edges never change; the layout is recomputed in the map's current Align
style, saved like Arrange, and Undo restores the previous placement. Arrange keeps the
chosen order afterwards. The last key is remembered per map only to highlight the menu
item — opening a map never rearranges it. Not shown in My map, for readers, in public
views or in kanban mode (rules-driven layout).
- **Bottlenecks button**: when switched on it shows "1 + 3" — real bottlenecks in red and
the potential (orange) ones that just lit up on the canvas. Switched off it keeps
reporting real ones only.
- **Rose skin** (12th built-in): rose pink with a wine primary, dark mode as deep wine,
destructive shifted to orange-red so it never blends with the primary. New `rose`
background pattern — a rose plant seen from the side (stem, thorns, leaves, bud) —
added to the pattern enum (validator, gallery and skin editor in sync).
- New suites: `usporadani-uzlu` (pure unit), `ui-usporadani`; `ui-hrdla-mapa` and
`ui-ikony-uzlu-kontrast` extended. Lite bundle cap 505 → 510 kB (skin data + pattern SVG only).
Upgrade notes: one migration (`1788721920_ruze_skin.js`) extends the `skin_id` /
`builtin_id` enums — additive, no data rewrite. No configuration change. Custom skins that
use the new `rose` pattern will not validate on instances older than this version.
## v0.58-beta — 2026-09-02
**Bottlenecks — honest edition: map badges, "Where it stalls most", My day warning; PDF reports**
- **Bottlenecks in the map editor**: goals that are overdue — or stalled for 14+ days while
holding up other unfinished steps — get a red "Bottleneck" badge that is always visible.
A new 🔥 toolbar toggle additionally highlights *potential* bottlenecks (a goal with ≥2
unfinished downstream steps, orange) and traces a dashed critical path to the real ones.
Facts only — days overdue, steps held, days without movement; no invented scores.
Stalling uses the same single definition as "Not moving" (the change log, 14 days).
- **Organization page**: new "Where it stalls most" section — real bottlenecks across team
and shared projects with a jump-to-map link; section order is now Projects by % → Where
it stalls most → Overdue → the rest. Included in the Report export (MD/CSV).
- **My day**: your own overdue goal that holds up other steps carries a discreet
"bottleneck · holding N steps" tag — the person who *is* the bottleneck learns it first.
- **Report → PDF**: two new export options — "PDF (to share)" (always light on white) and
"PDF (as I see it)" (your skin and dark mode). Fixes exports that came out dark-on-white
unreadable, half-recoloured cards, and white bands over dark page edges; the project
dashboard PDF benefits from the same fix.
- **Align button**: Compact and Bands styles now also take effect on maps whose categories
have just two sub-goals — previously all three styles produced identical layouts there
and the button appeared to do nothing.
- New endpoint `GET /api/kb/map-activity` (per-map stall stamps; same visibility rules as
map history — not available on public map views). New suites: `ui-hrdla-mapa`,
`ui-organizace-hrdla`, `zarovnani-dvojice`.
Upgrade notes: no migration, no configuration change. Do **not** set `KB_STUCK_DAYS`
in production — it silently changes the stall definition for the whole instance
(it exists for e2e tests only).
## v0.57-beta — 2026-09-02
**Timeline (Gantt) view on the Tasks page**
- **New “Timeline” view** (second tab: Table | Timeline | Board | Calendar): a horizontal
Gantt-style axis grouped by project — goals and tasks as bars, milestones as ◇, overdue
items with a red ring and a “Today” marker line.
- **Three scales** — Days / Weeks (ISO week numbers with date ranges) / Months (quarter +
month header) — plus −1/Today/+1 stepping, a jump-to-project picker, a scroller strip and
drag-to-pan (mouse and touch); a compact mobile layout for narrow screens.
- The map filter on the Tasks page no longer offers “No map” — every task lives in a project,
so the option only ever produced an empty list.
- The chosen view is remembered per browser (as before with Table/Board/Calendar).
- New test suite `ui-casova-osa.js` (red on the previous build — the tab does not exist).
Feature contributed via the Antigravity agent (reviewed, re-based onto the current
architecture and localised before merging).
## v0.56-beta — 2026-09-01
**Task cap decision (owner, 1 Sep): 500 with visible completion; anonymous export in the menu**
- **Tasks page:** loads the first 500 tasks via `listPage` (`-created_date,id` tiebreak) and, when
more exist, shows "Showing 500 of N tasks" with a Load-all button (pages of 500). Tasks over the
old 1000 cap used to vanish silently. Per-map queries unchanged; note: the shared hook means the
Home page also loads the first 500 (My day itself is computed server-side).
- **Editor:** "Export JSON (no names)" also in the ⋮ menu — it existed only in the wide toolbar.
- New suite `ui-strop-ukolu.js` (520 tasks; red on the previous build — no banner, silent cap) and
an extended `ui-sablona-z-mapy.js` (menu item + no-names download, red on the previous build).
## v0.55-beta — 2026-09-01
**Post-marathon polish: review-panel fixes, sharing guard, one toolbar definition**
- **Autosave:** leaving a map while a slow save is in flight no longer loses the follow-up edit —
the flush awaits the in-flight PATCH (promise ref, not the effect-owned timer) and sends the
current state with the returned `base_updated`. Proven red on the previous build.
- **Lowercase follow-up:** migration `users_email_lowercase_2` also rewrites `assignedBy`,
`holder`, `deputy`, `automationRequestedBy`, `deadlineChangeRequestedBy` inside map nodes
(paged, logged); a users update hook keeps the e-mail lowercase even through PocketBase's
confirm-email-change; the twins log now states truthfully that the older of a pair takes the
lowercase address (behaviour unchanged, owner-approved).
- **Sharing guard (server):** `syncShares` silently skips invalid share values (no `@`,
whitespace, `ext-…@kontakt.invalid` pseudo-addresses) with a warn log — a malformed value used
to abort the whole share sync, and an external-contact pseudo-address could gain a `work` row.
- **Editor:** Undo works after "Add goal"; the toolbar and ⋮ menu render from ONE action list
(byte-identical DOM; the list documents their historical asymmetries instead of hiding them).
- **API boundary:** `nodeStatus`, org-structure and member-admin calls go through `api/kb.js`.
- **Cloud:** one Stripe client (`stripe_klient.py`, url-encoded params), one SMTP sender with
PDF attachments and a CRLF guard; revenue prediction already fixed to local clock in v0.54's wave.
## v0.54-beta — 2026-09-01
**Autosave correctness (F1-01, F1-03) + Tasks page split into hooks (F3-10)**
- **F1-01:** the autosave effect now serialises its own PATCHes — while one is in flight the next
round is rescheduled (max one waiting) and sends the CURRENT state after it lands. Before, a save
slower than the 1.2 s debounce made the editor conflict with itself (409 → "map changed elsewhere",
"Reload" dropped the last edit); the draft branch could create two projects. The draft create no
longer sets `skipNextSave` — text typed during the flying create goes out with the next round.
- **F1-03:** the fingerprint stored on load is canonical (`cleanMap`), so an older record with fewer
node-data keys no longer triggers an empty PATCH on the first dimensions change. The merge base
stays raw — three-way merge unchanged.
- New suite `ui-autosave-serializace.js` (CDP-delayed PATCH response; raw record written straight to
SQLite): red 7/6 on the previous build, green 13/0 after; the PATCH ≤ 3 network anchor holds.
- **Tasks page (F3-10):** `Tasks.jsx` 1,148 → 679 lines, four hooks (`useTaskFilters`,
`useTaskTrees`, `useMapNodeActions` with real `useCallback` deps, `useTasksPageData`).
**Declared change:** `nodeTrees` = `rawTrees` (deps: maps) + prune (deps: filters) —
`computeWaitingSet` and tree building no longer rerun on every search keystroke; result identity
proven on 14 scenarios against the old implementation.
## v0.53-beta — 2026-09-01
**Lowercase e-mails + optimistic-lock row actions (debt 1+2 after v0.46; S4-01, S6-02, S3-04, S4-02)**
- **Server:** the users create hook lowercases the e-mail (API and OAuth registration). PocketBase's
unique index and all sharing/rights/My-day matching are exact — an account `Jan.Novak@…` never
saw a project shared to the lowercase address, and login was case-sensitive.
- **Migration `users_email_lowercase`:** rewrites `users.email` and every reference (owner_email,
created_by, deputy, assignee_email, triggered_by, invited_by, `map_shares.email`/`email_edit`,
author/actor fields, JSON `shared_with*` and `nodes[].data.owner` in maps) — schema-driven;
`clients`/`externi_kontakty` untouched. Twins (`Dup@` vs `dup@`) are NOT changed, only logged —
instance boot never fails on customer data. Idempotent.
- **Frontend:** login/reset/registration send lowercase; `lib/mapNodes.ulozDoMapy()` = fresh read +
`base_updated` + one retry on 409, used by add-to-map, task status patch and Tasks-page row
actions (server-side enforcement of `base_updated` stays planned for v1.0).
- New suite `emaily-lowercase.js` (15 checks incl. an upgrade over a volume from a pre-fix image;
red 5/6 against the old code). ⚠️ Two silent Goja pitfalls found by the upgrade test:
`field.type` is a method, JSON fields must be read via `record.getString()`.
## v0.52-beta — 2026-08-29
**Fourth wave from the code review — frontend structure (part 3): editor split complete**
- **`useMapAutosave`** (F1-07 step 13): the autosave effect (incl. the draft branch), merge base
(`zapamatujServer`, `zrcadliStavDoZakladny`), conflict/remote-change handling, `nasadNaPlatno`,
`slitCiziZmenu`, `handleKeepMine` and the background watcher moved verbatim (427 = 427 lines);
`skipNextSave`, `nodesNow/edgesNow`, `mapRulesNow`, the load effect and `handleSaveTemplate` stay
in the editor. Effect order unchanged.
- **Network anchor** in `ui-autosave-odchod.js`: a deterministic editor session (add sub-goal →
rename → autosave → Undo) must stay at ≤ 3 `PATCH` and 0 extra `GET`/`POST` — measured twice
before and after the move.
- **JSX sections** in `components/goal-map/editor/` (ConflictBanners, EditorToolbar, PersonalTabs,
TitleStrip, LeftRail, EditorDialogs) — presentational only, 897 JSX lines byte-identical after
re-inlining; `<ReactFlow>`, the selection bar and `BulkEditDialog` stay in the editor (context
providers). `GoalMapEditor.jsx` is now **1,517 lines** (was 3,717 before the wave).
- Gates: full click-test 60/60 and full regression 165/165 on the clean tree, twice (after step 13
and after the JSX split); per-step domain suites along the way.
- **Cloud admin:** the revenue prediction now uses the LOCAL clock like the rest of the overview
(`predikce_obratu` had drifted to UTC — between 00:00 and 02:00 CEST on a month boundary it
targeted the wrong month; caught by the nightly regression on 1 Sep).
## v0.51-beta — 2026-08-29
**API and MCP: unknown fields are rejected, and assistants get the plan (`planned_on`)**
- **Behaviour change (breaking for sloppy integrations):** every v1 write endpoint and every MCP
tool now answers **400 / -32602** to a field it does not know — top-level, inside `tree`/`items`
(recursively), inside rule `trigger`/`conditions`/`actions` — and the message lists the allowed
fields. Until now an unknown key was silently dropped and the call returned 200: an AI assistant
asked to "set priority high" reported success over an unchanged map. Common names from other
tools get a hint: `priority` → `planned_on`, `due_date` → `deadline`, `assignee` → `owner`,
`tags`/`labels` → map structure or `color`, `reminder` → a `deadline_approaching` rule,
`estimate` → not kept; camelCase → the snake_case name.
- **`planned_on` in the API and MCP** (`POST …/nodes/{nodeId}`, `tree`/`items`, `update_node`,
every read): *when the key owner plans to work on it*, today to 7 days ahead, empty string
clears — the same choice the app's row bar offers. This is how killBottleneck expresses
priority (no priority field, on purpose; the deadline is an agreement and stays put). A node
planned via the API lands in **My day** exactly like one planned in the app. A date outside the
window is a 400, not a silently ignored value. Only editors can set it through a key; readers
with their own work keep status-only.
- MCP: `additionalProperties: false` on all 17 tools in both servers (HTTP `/mcp` and the npm
stdio package — the stdio package used to *advertise* it while dropping the keys at runtime);
`get_map` shows `plan: YYYY-MM-DD` next to the deadline.
- Rules over v1/MCP: unknown keys inside `trigger`, `conditions[]`, `actions[]` and
`create_subnodes.items` are rejected; the rule builder in the app is unchanged.
- Tests: new `api-neznama-pole` suite (v1, MCP HTTP, MCP stdio, My day); `mcp-http` parity now
also checks `additionalProperties`.
**Upgrade notes:** no migration. If an integration of yours sent fields the API never
documented, it will now get a 400 with the list of allowed ones — fix the field name. The npm
package `killbottleneck-mcp` needs the matching release for `planned_on` and strict arguments.
## v0.50-beta — 2026-08-28
**Fourth wave from the code review — frontend structure (part 2): the map editor split into domains**
- **`GoalMapEditor.jsx` 3 717 → 2 575 lines** (F1-07 steps 1–12). Pure logic in `lib/personalMap.js`,
`lib/mapProgress.js`, `lib/nodePermissions.js`; domain hooks `useMapCounts`, `useMapHistory`,
`useMapExport`, `useMapRules`, `useAiActions`, `useMapLayoutRefs` + `useMapLayout`, `useBufferInsert`,
`usePersonalMapView`. Code moved verbatim (bodies, deps, comments — verified by script in both
directions); `contextValue` shape unchanged; `skipNextSave`, `nasadNaPlatno` and the load effect stay in
the editor (autosave = next sub-wave).
- **Fix (F1-05):** `pushHistory` reads the latest `nodesNow/edgesNow` refs instead of a closure — the
“Repair tree” toast held a stale handler and Undo after it returned the map without nodes added meanwhile.
- **F1-06:** `recenterMap` memoised (defined after `rfInstance` — deps above it would hit the TDZ), so the
align-lock effect and `handleAlign` no longer recompute every render.
- **Effect order note (step 12):** three layout effects (`alignMapKey`, `alignLock` from the account, style
lock/cleanup) now run before the archive-offer effect and the rules load; they share no state or refs with
them. `useMapDirection`'s matchMedia/cleanup effects run after the `org` effect; `pendingDeepLink` before
the load effect.
- New unit test `personal-map.js` (41 checks) + `tests/_alias-loader.mjs` (node hook for `@/` imports).
- Gates: full click-test 60/60 and full regression 165/165 on the clean tree.
## v0.49-beta — 2026-08-28
**Fourth wave from the code review — frontend structure (part 1)**
- **Instance config:** one shared loader (`hooks/useKbConfig.js`, shared in-flight promise,
in-memory only, invalidated after writes that change it — registration, purpose, AI settings —
and forgotten on login/logout). Home used to fire up to six `GET /api/kb/config`; now one.
- **One boundary for `/api/kb/*` calls:** `api/kb.js` (`kbSend` with client-side timeout, no axios-like
`{ data }` wrapper); `functions/` folder removed; one error convention (`err.response?.error`).
- **Project creation:** `createProjectRecord` is the single path (empty project, template from the
dialog, AI preview, "Use template" from the editor preview, draft autosave) — five hand-built
`GoalMap.create` bodies before. Empty/AI projects still share with nobody.
- **Exports:** shared core in `lib/saveFile.js` (`downloadText`, `csvEscape`, `savePdf`,
`safeFilename`, `dateStamp`, `afterRepaint`). File-name policies preserved 1:1.
**Behaviour change:** the date stamp in task CSV/MD, full data export and My-day PNG file names
is now the local day (was UTC → yesterday after 22:00 CEST); CSV now quotes a lone `\r` too.
A map **without a title** exported to PNG/PDF is now named `mapa-cilu.png` / `goal-map.pdf`
(the fallback used to go through the same character strip as titles → `mapacilu`); named maps unchanged.
- **Dialogs:** `useDialogForm` + `BusyIcon` in 13 form dialogs (busy flag under six names, double-submit
guard, error, close, Enter). Side effect: creating an API key with Enter during save no longer
creates it twice.
- **Tasks page:** `useSidePanels` replaces three copies of the buffer/time-log toggle;
`TaskTable.jsx` 840 → 352 lines — rows in `components/tasks/table/`, handlers via
`useTaskTable()` context instead of 32 drilled props (component bodies moved verbatim).
- **Map editor (first steps of the domain split, F1-07):** pure functions out of
`GoalMapEditor.jsx` into `lib/personalMap.js`, `lib/mapProgress.js`, `lib/nodePermissions.js`
(bodies moved verbatim, new unit test `personal-map.js` with 41 checks) and three hooks
`useMapCounts`, `useMapHistory`, `useMapExport` — editor 3 717 → 3 259 lines; `contextValue`
shape unchanged.
- **Lite mode stays at 500 kB:** `kbSend` core and the My-day/Organisation reads live in small
modules (`api/kbSend.js`, `api/myDay.js`) so the lite bundle does not drag the whole `api/kb.js`
into its shared chunk; `api/kb.js` re-exports them.
## v0.48-beta — 2026-08-27
**Third wave from the code review — test foundation, one source for system templates, dead code out**
- **Behaviour change (approved drift fix):** a person assigned in a template gets the project as a
*collaborator* (`work`) when the project is created from the app, exactly as the automatic
creation already did — until now the app granted `edit` (decision of 7 Aug 2026, S5-03).
- **System templates:** the 40 built-in templates now come from one file
(`pb_migrations/data/system_templates.json`) and one migration instead of six layered
migrations (6 369 lines). Existing instances are untouched (every template already exists);
fresh instances get the final state directly. Dedup is by title *and* `owner = ''`, so a
user's own template with the same name no longer blocks a system one.
- **Stripe:** period end is read with a fallback to the subscription item (API `basil`).
- **Tests:** shared `product/tests/_harness.js` (docker-assigned ports, image-hashed container
names, mandatory `KB_TEST_IMAGE`, one summary format, `HARNESS_MUTACE` self-test); five suites
converted; `tests/run-all.sh` discovers suites itself (two suites had been missing from the
hard-coded list three times).
- Removed the dead Base44 export `app/` and `sync/sync-from-base44.sh` (private monorepo only).
- Cloud: one `.env` loader (`konfig.nacti_env`) instead of eight copies.
**Upgrade notes:** one automatic migration (system templates seed — no-op on existing data).
No configuration changes.
## v0.47-beta — 2026-08-27
**Second bug-fix wave from the full code review — write ordering, one core for rules, faster overviews**
- **Editor:** editing and leaving the map within 1.2 s (Back, logo, another map opened from inside the
editor) now flushes the pending save instead of dropping it; Undo after Align is saved.
- **Notifications:** e-mail-only recipients no longer get deadline reminders again after every server
restart (dedup barrier now also covers the e-mail path).
- **Agents:** two concurrent callbacks for the same run — exactly one wins (atomic token claim); the run is
closed only after the map write, so a failed write marks the run `failed` instead of `done` over an open
node; a callback arriving before the webhook 2xx no longer revives the token; the dispatch cron cannot
overlap itself; rule-chain depth is carried through agent runs (`agent_runs.depth`).
- **Rules:** one shared core (`rules-api.js`) for the app routes and the v1 API. **Behaviour change:** a full
edit via `POST /api/kb/rules/save` without `enabled` no longer silently re-enables a disabled rule (the v1
semantics apply everywhere; the app UI always sends `enabled`). A failed scheduled run no longer burns
its dedup key, so a fixed rule fires again.
- **Sharing / import:** map and its share rows are saved in one transaction; creating a map no longer
fails on share sync; `/import-all` skips a map whose nodes exceed 5 MB with a reason and imports the rest.
- **Trial lock:** read-only POSTs pass after the trial ends — MCP `initialize`/`ping`/`tools/list` and the
read tools, and `share {action:"list"}`; writes stay 402.
- **MCP over HTTP:** argument types and enums are validated like the stdio server — `"false"` is rejected
with `-32602` instead of being coerced to `true`. **Behaviour change** for HTTP clients that sent strings.
- **OAuth (MCP connectors):** connecting again deletes your own *expired* API keys first, so the 20-key cap
no longer blocks reconnecting after months of use. **Behaviour change:** the key list may shrink on its own.
- **Access checks:** six map-access helpers collapsed into one computation; My day and the full export look
up shares in one batch (≈520 → ≈10 queries for 500 shared maps); `jeAdmin()` replaces 31 hand-written role checks.
- Tests: new suites `notify-email-dedup` and `ui-autosave-odchod`; concurrency, type, OAuth, import and
trial checks added; an align-order test made deterministic.
**Upgrade notes:** two automatic migrations (`agent_runs.depth`, `mail_budget.day` max 250). Integrations:
see the three behaviour changes above (rules `enabled`, HTTP MCP types, OAuth key cleanup) and note that
an agent run is reported `failed` when the node could not be marked done. No configuration changes.
## v0.46.1-beta — 2026-08-27
**Bug-fix release from the full code review (wave A) — no new features**
- **Organization → Report → Markdown works again.** Since v0.44 it threw `md is not a function`
for any organization with at least one project — no file, no message. A UI test now clicks it.
- **My account: the display name no longer disappears after a reload** (the user DTO dropped
`name`, so the next save wiped it).
- **Daily e-mail digest is sent once after a server restart**, not twice: the "sent today" mark
was 17 characters long but the field allowed 10, so it was never saved (migration widens the field).
- **Sharing validates the e-mail address first.** A typo without `@` used to be written after the
map was saved, poisoning the map's share list (later shares failed, removing one member dropped all).
- **Orphaned attachments are cleaned up beyond the first 500 files**; the nightly job now pages.
- **Creating a project no longer fails silently** (expired trial 402, seat limit 409, network).
- **API keys: `token_hash` is now a hidden field** — the raw collection list returned it to its
owner; the test that should have caught it could not fail (`|| true`). Two more tests fixed the same way.
- Housekeeping: 32 unused npm packages removed (node_modules −32 MB), ESLint now covers the whole
frontend `src/` (44 % of files were unlinted), 5 dead UI files and 25 unused translation keys removed,
missing translation keys added (a raw key showed in the org-position dialog and in AI error toasts),
EN welcome map link fixed (404), `@capacitor/*` moved to runtime dependencies.
**Upgrade notes:** two automatic migrations (`mail_budget.day` max length, `api_keys.token_hash`
hidden). No config changes.
## v0.46-beta — 2026-08-27
**An API key acts as its owner — shared and team maps through API and MCP, plus `get_portfolio`**
- **API keys and MCP now see and edit exactly what the key owner can in the app**: own maps,
team maps and maps shared with the owner. The share level decides what a write may do —
`owner`/`edit` = full write; `work` (collaborate) and `read` = only the `status` of the
owner's own nodes (like ticking off in the app), other fields 403. Rules and their run log
need edit rights (as in the app). `GET /v1/maps` and
`GET /v1/maps/{id}` return the level as `access`; `list_maps`/`get_map` show it too.
- **Assigning an owner through the API shares the map with that person** as a collaborator
(`work`), exactly like the app does — the assignee finally sees the work in My day. Never
downgrades an existing share, never for external contacts, only when the key owner may share
(map owner or named editor); the response says who was shared with in `shared`.
- **New MCP tool `get_portfolio`** (+ `GET /api/kb/v1/portfolio`): the Organization page for
assistants — completion per project, overdue and stuck items, people with overdue work,
changes in the last 7 days — over the team and shared maps the key owner can read.
- Still true: the role is never read (an admin's key sees no one's private map), a `read`-scope
key never writes, someone else's private **and public** maps are 404, the org map stays
read-only through a key, administration/AI settings/users are never reachable.
**Upgrade notes (breaking for API/MCP integrations):** a key now reaches **more** than before —
shared and team maps used to be deliberately unreachable (404). If an integration relied on
"a key sees only its owner's maps", review it: `list_maps` may return more maps, and writes on
them follow the share level (403 where the owner is only a reader or collaborator). No migration.
The npm package `killbottleneck-mcp` 0.46.0 adds `get_portfolio` — and it is the first npm
release since 0.35.0, so `npx killbottleneck-mcp@latest` also picks up everything from 0.36–0.45
(rule tools, `get_org_structure`, `list_people`). Older clients keep working (they simply do
not offer the new tools).
## v0.45-beta — 2026-08-26
**Download all my data — leaving is part of the product**
- **My account** has **Download all my data**: one JSON file (`killbottleneck.export/1`)
with every project you can see — each in the same shape as a single map export, so it
imports elsewhere — plus tasks, rules, comments, the change log, the attachment list,
who can see the project, the idea stash, time tracking, external contacts, notifications,
rule templates and the member list.
- It works even after the trial has expired: the “Download data” link sits right in the
top bar.
- **Upload data from an export** brings a whole file back: every project (with rules,
archived ones stay archived) and the idea stash — into the same or another instance.
- Nothing you cannot see leaves with you: other people's private projects and public
notice boards are not in the file, members come as a safe subset (no secrets).
**Upgrade notes:** no migration. New session endpoints `GET /api/kb/export` (5 per minute,
one at a time per instance) and `POST /api/kb/import-all` (50 MB, 2 per minute). Very large instances are truncated per list and the file says
so in `truncated`.
## v0.44-beta — 2026-08-25
**Organization: the view from above for admins and managers**
- The top bar gained **Organization** — admins and managers see on one screen what
is overdue across projects (who and for how many days), how far projects are,
what has not moved for 14+ days and who has the biggest backlog.
- It counts only team and shared projects — a private project is never counted,
not even in the totals, and the page says what it counted.
- The Report button downloads the same thing as Markdown (Monday report) or CSV
with the numbers you see on screen; at the bottom there is “What changed in the
last 7 days” across projects.
- Clicking an item jumps straight to the goal in the map; clicking a person opens
Tasks pre-filtered to them (`/tasks?assignee=<e-mail>`).
- Fixed: switching the language in the account menu failed after opening
Organization settings or billing.
**Upgrade notes:** no migration. New session endpoint `GET /api/kb/portfolio`
(admin and manager only, 403 otherwise). “Not moving” (for goals) and “What
changed” read the change log, so on maps untouched since the log was introduced
they may start out empty; tasks are judged by their last change right away. Not
available through API keys / MCP yet (the key still acts only on its owner's maps).
## v0.43-beta — 2026-08-25
**A starter map without deadlines, a “what is it for” question and two projects to begin with**
- The starter map no longer scares anyone with deadlines — tour items carry only
a plan (“I want to do this”): they light up in My Day for the first days, never
turn red, and whatever you skip stays in the map.
- On the first login the first admin is asked once: “What will you use
killBottleneck for?” — company or team, family and friends, or just yourself —
and the starter map adapts (a solo user is not told to “assign roles”).
- Every new account gets two projects: the starter map and a small trial project
for the chosen purpose (A better working day · Shared joy · Treat yourself), so
My map makes sense right away.
- The instance purpose can be changed any time in Organization settings — it
applies to newly invited people, existing maps stay as they are; invited people
inherit it and never see the question.
- On the phone your own entries sort above the tour items.
**Upgrade notes:** the `org_settings.purpose` migration runs automatically. On an
instance with a single admin the purpose question shows once after the upgrade —
it only affects newly invited people, existing maps stay. `KB_PURPOSE_ASK=0`
disables the question.
## v0.42-beta — 2026-08-25
**See what is stuck at others; the agent assigns only to real people**
- The My Day panel gained an “Overdue at others” number — see at a glance
how much of the work you delegated is already late; click opens the list.
- Whoever loses a goal or gets it handed to someone else is now notified —
a silent move no longer surprises anyone.
- The AI agent can list the people of the instance (list_people tool) and
assigns work only to a real member or contact — a typo in an e-mail is
rejected with a hint of who you probably meant.
- The API keys dialog shows the instance address and a ready-made command
to connect Claude Code; the copy buttons work over plain http too.
- The three AI advisor questions before generating goals are no longer
mandatory.
- The project dashboard counts tasks as goals with an owner or a deadline —
it no longer reports “no tasks yet” for a project that has them.
- Two instances side by side on one host: the container name can be set
with KB_NAME.
**Upgrade notes (breaking for API/MCP integrations):**
- `owner` in `POST /v1/maps`, `POST /v1/maps/{id}/nodes` and `update_node` must be the e-mail of an
instance member (or a visible external contact). Unknown e-mails now return **400** with a hint
instead of being stored silently — use the new `list_people` tool / `GET /v1/members` first.
- `scope` when creating an API key must be `read` or `read_write`; anything else is **400**
(previously silently downgraded to `read`).
- New notification type `node_unassigned` (migration `1787400000`); `KB_NAME` lets you name the
container. The npm package `killbottleneck-mcp` 0.42.0 ships with a later release — until then
`list_people` is available through the built-in HTTP MCP endpoint (`/mcp`).
## v0.41.2-beta — 2026-08-25
**Encrypted backups and security updates for bundled libraries**
- Data backups can now be encrypted with a passphrase — set KB_BACKUP_PASSPHRASE
when backing up and nobody can read the archive without it (GPG, AES-256).
- Restore handles encrypted as well as older plain backups — nothing to convert,
and a forged archive is refused before it touches live data.
- Security updates for the bundled libraries — 7 reported dependency
vulnerabilities fixed (dompurify and build tooling among them).
**Upgrade notes:** nothing to do — encryption is optional; without KB_BACKUP_PASSPHRASE backups behave exactly as before.
## v0.41.1-beta — 2026-08-24
**Screenshots in bug reports, a reliable day star, and timer start from the panel**
- Bug reports and ideas can now carry a screenshot — just paste it with Ctrl+V
right into the text, or pick a file. The image is scaled down automatically,
arrives as a mail attachment and shows up in your "Already reported" list.
- The "top today/tomorrow" star switches correctly when you move the same task
to the other day; clearing removes it everywhere and can be undone in one click.
- Time tracking can be started right from the left panel — an empty panel is no
longer a dead end.
- Entries in the time-tracking panel got their own background, so task names and
the assignment row no longer blend into their surroundings.
## v0.41-beta — 2026-08-21
**What's new:**
- The "Edit" level is a co-manager: whoever hands out work on a map can also share it
with more people. Team access, public link and map deletion stay with the owner.
- Anyone who was given work can request a different due date on their own step —
even with view-only access.
- The sharing list tells the whole truth: a viewer with work shows "has work here",
and team maps got a section for people who have work via team access.
- Sharing levels now have descriptions — you can see what each level adds.
- External contacts stand out on the map and in lists: a name badge with "(external)",
so a note about a partner doesn't look like work someone is doing.
- An access upgrade notifies the recipient; assigning work doesn't duplicate notifications.
- Fixed: buttons on step cards are clickable with the mouse in read-only mode
(broken since v0.20).
## v0.40-beta — 2026-08-20
**Connect the app to OpenAI, OpenRouter or any other service with an API key**
- **A new AI mode: `openai`.** Until now killBottleneck could only speak Ollama's dialect,
so an ordinary API key — OpenAI, OpenRouter, Groq, Mistral, Together, or your own vLLM,
LM Studio, llama.cpp or liteLLM proxy — got you nowhere. Now you enter an address, a key
and a model name and every AI feature works. Reported from the beta.
In the app it is **Administration → AI features → OpenAI-compatible**; in `.env` it is
`KB_AI_PROVIDER=openai`. **Test connection** tells you straight away whether the key is
valid and whether the model name exists.
- **Dictation works through the same service.** No separate transcription endpoint needed;
the model is `whisper-1` unless you change `KB_AI_TRANSCRIBE_MODEL`. If you would rather
run transcription somewhere else, `KB_AI_TRANSCRIBE_URL` still wins.
- **"OpenAI-compatible" is a family, not one interface**, so the app copes with the
differences: a service that does not accept a structured-JSON request is asked again, more
simply. A reasoning model that burns its whole budget on thinking and returns nothing now
**says so** instead of silently doing nothing.
- **The contract for `custom` is finally written down** — the "your own endpoint" mode has
existed for a long time, but what such an endpoint must do lived only in our source code.
New reference page: *Custom AI endpoint*.
- Nothing changes for existing instances: `ollama`, `api` and `custom` behave exactly as
before, and AI stays off by default.
**Upgrade notes.** Two migrations: one adds a fifth value to the AI switch, the other adds a
transcription-model field. Nothing is rewritten and nothing is switched on by itself.
⚠️ **Going back to v0.39 with `provider=openai` saved turns AI off silently** — the older
version does not know the value and answers "AI is disabled". Switch the provider back before
downgrading. With `openai` there is now also an hourly cap of AI operations per person
(`KB_AI_MAX_PER_HOUR`, 60 by default), because every call spends your own credit.
---
## v0.39-beta — 2026-08-19
**See what happened to a goal, edit a whole selection at once, and read the map from the lines**
- **Every goal now has a History.** A category in the goal dialog lists what happened to
it — one row per event, **with the date and the time**, newest first: status, deadline,
owner and name changes, moves under a different parent, changes to the brief, icon,
colour and performer, plus comments and attachments added and automation rules that
fired. **A rule shows up as a rule**, not as the person who once wrote it; until now the
recorder stored the rule author's address, so the log claimed a human had clicked.
Comments and attachments are read from where they live, so the history is complete
**retroactively** — you get it for goals you created months ago.
The history says *that* the brief or a comment changed, never *what* it said: it is a
record of movement, not a second copy of your data. It does not leave a publicly shared
map, it reaches 400 days back, and moving a card on the canvas is not recorded (tidying
the map would bury everything that matters).
- **A selection can be edited in one go, not just deleted.** Shift-drag several goals and
set status, owner, deadline, icon or colour for all of them. Every field has its own
switch, so only what you switch on is changed — "set the owner" never wipes a deadline —
and a switched-on field left empty **clears** the value, which is how you strip an owner
or a deadline in bulk. A deadline somebody else set is skipped and the dialog says so up
front; without that the server would refuse the whole save and nothing would happen.
Undo reverts the lot.
- **The connector lines now carry the state of the goal they point to.** Green and still
means done; red and moving faster means past the deadline; everything else looks exactly
as before. A done goal with a missed deadline still counts as done. The two new colours
are **part of the skin**, so they suit every theme — in Ruby, where the lines are red
anyway, "past the deadline" shows up as flame orange. With animations switched off
system-wide you get the colours without the motion.
- **The goal dialog shows where things are.** "Attachments" and "Tasks & comments" carry a
count, so a comment you can see on the card is no longer something you have to hunt for.
- Selecting a line makes it thicker rather than recolouring it — the colour now means the
state of the goal, and two different things in one colour could not be told apart.
**Upgrade notes**
- One migration adds the new event types and a field recording *what* made the change.
Nothing is rewritten and no existing history is touched.
- Skins keep working unchanged. A skin without the two new colours is still valid and
falls back to the built-in green and red, token by token.
---
## v0.38.1-beta — 2026-08-19
**Bug reports without your address or your company name**
- **Bug reports are sent anonymously.** Your text, the app version, the page you were
on and your browser go out — your e-mail address and the instance address do not.
Bugs get fixed in the program for everyone, not on individual accounts.
- **Want a reply? Tick the box.** Only then is your address attached and put into
`Reply-To`. Without the tick it goes nowhere.
- **Sent reports delete themselves after 30 days.** Until now `reports` was the only
collection in the app with no clean-up at all.
- Added a guide to bug reporting (Czech and English), documented `KB_REPORT_TO` in the
environment reference, and lined the privacy policy up with what actually happens.
---
## v0.38-beta — 2026-08-19
**Describe a process, not just name it — and tell us when something breaks**
- **The goal description now takes formatting.** A toolbar above the field offers
bold, italic, strikethrough, two heading levels, bulleted and numbered lists, and
links; Ctrl+B and Ctrl+I work. A Preview switch shows the result. People asked for
this because they document whole processes in there, and a plain box was not enough.
The text is stored **as markup, not HTML** — every existing description stays valid
and nothing is migrated.
- **Links in the description can carry a name.** The link button offers the goal's own
**attachments**, so the description reads "evidence" instead of a full-width Google
Sheets address. You can also type any address and name it yourself.
On the map card the description still shows as plain text — a card is one or two
lines and markup would look like a defect there.
- **Icons for goals: about 200 of them**, in categories, with search in Czech and
English (accents optional), and a field for any emoji from your keyboard. The ones
you pick most recently stay at the top under Favourites. The catalogue loads only
when you open the picker, so the app is not any heavier.
- **Hover help on the map icons.** Only some of them had it: the pencil in a goal's
footer, "Add sub-goal", the deadline badge, the comment bubble and the sticky-note
colour dots now say what they do.
- **A goal card shows how many attachments it has** — a paperclip badge next to the
comment bubble. Until now an attachment was only visible after opening the goal.
- **When a new version arrives, the bell says what changed** — a few sentences in
your language, right in the app. No link to an English changelog.
- **Report a bug or an idea** straight from the app — under the person icon and from
the left rail on the overview, in Tasks and in a map. The
message goes to our team and we can reply to you directly. Before you send, the
dialog shows exactly what travels with it: your address, the instance and the page
you were on. Nothing is collected quietly. **Self-hosted instances do not offer
this** — they have nowhere to send it and must not send anything out on their own.
---
## v0.37-beta — 2026-08-18
**Your organization everywhere you look — and the project name finally has room**
- **The project name moved out of the map toolbar and onto its own line** above the
canvas, in larger type. Squeezed between the icons, a longer name was cut off
mid-word. Click it to rename, Enter confirms. In its resting state it is text, not
an input: the old transparent field spanned the canvas and swallowed the mouse, so
a strip 960 px wide could neither grab a node nor pan the map.
- **The map toolbar now shows your organization's logo**, the same as the app header
does. Until you upload one, the killBottleneck mark stands there — one or the
other, never both. **Anonymous visitors of a publicly shared map always see the
killBottleneck mark**, not your logo: a public map is your calling card outward,
not an internal screen.
- **The browser window title starts with your organization**: "Acme killBottleneck".
With several windows open you can tell yours apart in the taskbar at a glance.
- **Project cards show the main goal** — the text of the apex node — under the project
name. The name tends to be shorthand ("FMEA — kanban") while the main goal is the
sentence that says what it is about. When the two match, the line is left out.
- **The simplified view shows the name day** in its header, the way the My day panel
in the full app always has (Czech version, on days that have one).
- **Fix: "Save as template" crashed the screen.** In the cloud beta, Export → Save as
template turned the whole screen black: the dialog called a state setter that an
earlier change had removed. Our linter stayed silent about it, because the ESLint
config pulled the recommended rule sets in a place where the `rules` key below
overwrote them — so not a single recommended rule was running, `no-undef` included.
Both are fixed, and a browser test now clicks the whole path.
- **Security: SVG is no longer an accepted logo format** (PNG, JPG and WebP are).
Uploaded files are served from the same origin as the app; an SVG opened directly
in the address bar can run script and reach the login token. Only administrators
can upload a logo, so this is a trap for the administrator, not a way in for a
regular user. Logos already uploaded keep working.
## v0.36-beta — 2026-08-18
**A structure manager: HR draws the org chart without being an administrator**
Drawing the company's org structure required the administrator role — which also
grants power over accounts, roles and instance settings. Handing that to an HR
person to let them maintain a position tree was far too much.
- **New flag: Structure manager**, granted per person in Organization settings,
independent of the role (the same pattern as the AI manager). The holder draws
the structure, appoints people to positions and sets deputies — and, because that
is the same job, may invite new people (always as members) and reset their
passwords.
- **What they deliberately cannot do**: change roles, grant any manager flag, delete
or archive the structure, publish it or share it out, or write to it through an API
key. Passwords of administrators and of other flag holders are off limits — a
password reset is an account takeover, and the boundary now covers flags, not just
roles.
- **They see only what they need**: Organization settings show them the list of
people and the structure, nothing else — no billing, membership, AI or instance
appearance, and no login history or map counts.
- **Nobody appointed? The administrator covers it**, exactly as before. Withdrawing
the flag takes every right away at once, including access to the structure map.
- Being appointed **arrives as a notification**, pointing to where the structure lives.
Vacated positions after someone leaves are now reported to structure managers too,
not only to administrators.
- **Fix: the Manager role description was a lie.** It promised "sees and manages all
tasks" — a rule removed back on 6 August. A manager's only remaining privilege is
inviting new people into the organization; the Member description was understated
in the same way (members create projects and invite colleagues just fine). Texts in
the app and the documentation now say what the code does.
## v0.35.3-beta — 2026-08-18
**The template preview is a demo — and your project is born clean**
Opening a template drops you into a preview where nothing is saved. People naturally
try things there — flip a card to Done to see the kanban move. The project was then
created from *that* clicked-around state: the card was born finished, no rule had
existed yet to move it, and the board looked dead on arrival.
- **A project from the preview is always created from the clean template.** Click
around all you like; none of it carries over. The one exception is the **name** —
rename the template in the preview and your project keeps that name (two projects
with the same title help nobody).
- **The preview bar says so plainly**: "Template preview — nothing is saved". The
preview deliberately stays unlocked; a demo you cannot touch teaches nothing.
- **Fix (silent data loss): leaving the preview for a real map switched off saving.**
Going straight from the preview to another map (avatar menu → Organizational
structure) kept the preview flag alive, because the route does not remount the
editor. Everything you then did on that real map was discarded without a word, and
a stray "Use template" button hung over someone else's map. Both are gone, and a
new test suite reproduces the loss on the old build.
- **A project created from the preview now shares like the dialog does**: people
assigned to nodes in the template get edit access and their assignment
notification. Previously only the "New project → From template" path did that.
- **If a template's automation rules fail to be created, you are told.** The project
used to be reported as fully created while its kanban was dead.
## v0.35.2-beta — 2026-08-18
**An invitation that no longer looks like spam**
Invited colleagues could not tell the mail was sent by a person they know, so some
reported it as spam. And once they were inside, closed the browser and came back a
week later, they had no idea what their organization was called or where to log in.
- **The subject line now starts with the address of whoever invited you** —
`richard@example.com invites you to killBottleneck — organization tengo`.
- **You can reply to an invitation.** It carries a `Reply-To` back to the inviter,
so anyone unsure can hit Reply and ask a person instead of a no-reply mailbox.
The footer says so instead of "do not reply".
- **"How to get back" moved into a card below the button.** Organization, sign-in
address and the e-mail you sign in with used to sit in a paragraph above the
button, where nobody read them as something to keep — and they competed with the
one thing you are meant to do right away: set a password.
- **New welcome mail after your first sign-in.** It arrives once the account
actually works and holds nothing but the way back: the address of your
organization and a nudge to bookmark it (Ctrl+D / ⌘+D). Only invited users get
it, and only once — guarded by a stored flag, not by guessing whether this is a
first login.
## v0.35.1-beta — 2026-08-17
**The kanban board stops crying wolf**
Marking a card Done moved it on the server, but the editor treated the rule's own
work as somebody else's change and showed the amber "someone else changed this map"
bar. The board looked broken while it was working perfectly.
- **An automation's change now merges into your unsaved work silently.** The card
slides to its new column in front of you and your half-typed edits stay exactly
where they were. Nothing interrupts you.
- **The bar still appears where it belongs**: when two people genuinely touch the
same goal, and when the change did not come from a rule. Your own work is never
silently discarded — when the editor cannot be sure, it asks.
- Why it only bit hosted users: the old code adopted the server's version only if
you had typed nothing since the save left. Locally the answer comes back before
anyone can type; over a real network you are almost always mid-sentence. Same
code, different latency — which is why the new test suite runs against a live
instance as well as a simulated slow link.
- MCP server listing in the Glama catalog (`glama.json` in the repository root).
## v0.35-beta — 2026-08-17
**Recurrence is back — on goals, powered by rules**
v0.34 removed task items and with them recurrence; this release brings recurrence
back the systemic way: as a property of goals, built on the automation rule engine.
- **A goal can repeat** — daily, weekly or monthly. Set it with the new
**Recurrence** switch in the goal detail (Assignment category). When the goal
is marked Done it returns to To do by itself and the deadline advances.
- **The rhythm is anchored to the original deadline**: every Monday stays
a Monday, the 31st stays the 31st (clamped to the last day in shorter months),
and missed occurrences skip to the nearest future one — a late completion
never hides how late you were, and never breaks the rhythm.
- **No new machinery**: the switch manages an ordinary automation rule
(`on Done → set_status todo + set_deadline advance`) visible in the map's
Rules. Hand-edit it and the switch honestly steps aside. Recurring goals
carry a 🔁 badge; templates and the API/MCP (`create_rule` with the new
`set_deadline.advance: daily|weekly|monthly`) transfer recurrence for free.
- **Fix: a node born straight into Done now fires status-change rules.** On
hosted instances, a quick "add subgoal → mark Done" could land in a single
save; the node was new in the diff, the kanban move rule stayed silent and
the card never left its column. Rules now treat a node born with a
non-default status as a status change (born as To do is not one). This also
applies when you paste or import a whole branch of finished nodes — each
fires the rule, with the existing cap of 10 rule executions per save
(anything beyond is openly logged as skipped).
## v0.34-beta — 2026-08-17
**One vocabulary: a task is a goal with an assignee or a deadline**
Words used to disagree across the app — the same box in the map was a "goal" in the
editor, a "task" in the table and a "node" in the API, while a second, separate kind of
"task" lived inside nodes. That second kind is now gone.
- **A task is a node (goal) with an assignee or a deadline.** New work = a new goal.
Nothing exists outside a map — quick thoughts go to the idea stash.
- **Standalone task items were removed.** They can no longer be created anywhere — app,
API or MCP. A migration deletes existing items and their comments; **time tracked on an
item is preserved** (re-attached to the item's node). The orange badge remains only as
a leftover-data detector: if you ever see it, something snuck in that shouldn't exist —
open it and delete the leftovers.
- **"New task" on the Tasks page now creates a goal** — under the project's main goal or
under a goal you pick — and immediately opens its detail to set the assignee and deadline.
- **⚠️ Breaking (beta): the `/v1/tasks` endpoints return 410 Gone** and the MCP tools
`list_tasks`/`add_task`/`update_task` were removed. Use `/v1/maps/{id}/nodes`
(MCP `add_nodes`, `update_node`) — a node with an assignee or deadline IS the task.
- Map import no longer creates task items (they are counted in `tasks_skipped`); templates
no longer carry `task_seeds` — assignees and relative deadlines live on the template's
nodes and keep working.
- **Removed with the items** (deliberately): task recurrence (a future feature will
revisit repetition on goals), item subtasks and item comment threads.
- Wording unified in Czech UI and docs: the responsible person is **"řešitel"**
everywhere ("garant" is gone); the person who assigned the task remains **"zadavatel"**.
**Upgrade notes.** If an integration of yours calls `/v1/tasks*` or the removed MCP task
tools, switch it to nodes: create work with `add_nodes` (set `owner` and/or `deadline`),
complete it with `update_node` → `status: done`. The migration deletes all task items and
their comments irreversibly — export anything you want to keep before upgrading.
## v0.33.2-beta — 2026-08-17
**"Create subgoals" works every time, and two ways to feed the map from a spreadsheet**
- **Fix: the automation action "Create subgoals" only ever worked on the first run.**
The second card (second complaint, second part) failed with a duplicate-id error,
the run was marked failed and the rule itself looked broken — the most appealing
piece of automation handled one case and then quietly gave up. Each run now gets
its own node id prefix.
- **New guide — Google Sheets integration**: every new row in the sheet creates a goal
in the map and unfolds the whole procedure under it (worked through on the 8D
report). Ready-made Apps Script you configure by filling in three lines, status
written back to the sheet, and honest limits at the end.
- **New guide — n8n integration** as a separate route: two ready workflows to download
(`.json`), import, fill in four lines. Unlike Apps Script it also reaches an
instance inside a company network, because it calls outward.
- Both guides cover the two shapes a map can take — the classic tree and the kanban
board where a complaint travels as a card through columns D1–D8.
- Installation docs now hold your hand outside Linux too: step by step for Windows
(Docker Desktop, PowerShell), Linux and macOS, including what actually trips people
up (a sleeping computer is a sleeping instance, access from a phone, the firewall).
## v0.33.1-beta — 2026-08-16
**Small things the first real installation turned up**
- **The version check no longer logs a 404** in the browser console. It asked GitHub
for the "latest release", which returns 404 for a project that only has a beta —
the behaviour was right (nothing was offered) but it looked like a broken app. It
now reads the list of releases and filters pre-releases itself.
- **New switch `KB_UPDATE_PRERELEASE=1`**: if you run a beta, you can opt in to being
told about the next beta. The default does not change — without it, pre-releases
are never offered to anyone.
- **The MCP server runs in Docker** (`mcp/Dockerfile`) and no longer exits when it is
not configured: it starts, offers its tools, and only a tool call tells you what is
missing. Clients like Claude Desktop no longer show it as broken.
- **The MCP server is on npm as `killbottleneck-mcp`** — `npx -y killbottleneck-mcp`
is enough, no need to clone the repository. It is also listed in the official MCP
server registry as `com.killbottleneck/killbottleneck`.
## v0.33-beta — 2026-08-15
**First public release — the self-host beta**
- killBottleneck goes public: this repository is the first public snapshot, released as a
**beta** (pre-release). The product is feature-complete and in beta — cloud and self-host
alike, one and the same app; what this repository tests is the self-hosted side —
installation, reverse proxies, SMTP, upgrades. Bugs → Issues, ideas → Discussions.
- The in-app option to order AI services from us was removed. Self-hosted AI means your own
model over [Ollama](https://ollama.com), or any compatible remote endpoint you configure
with an address and a token (`KB_AI_PROVIDER=api`/`custom`).
- The version check understands pre-releases: a beta install will be offered the final
release of the same number, and pre-releases are never offered to anyone as updates.
## v0.32 — 2026-08-15
**Kanban templates, and rules that travel with the map**
- Templates with built-in rules: a map template can carry automation rules. On project
creation the references are remapped to the new nodes and the rules are created the normal
way (validation and the 50-per-map cap apply, the creator becomes their author). Works from
every path — the New project dialog, the template gallery and automatic weekly/monthly
creation.
- New templates **8D Report — Kanban** and **FMEA — Kanban** next to the classic versions:
create cards under the first column and a finished card travels to the next step by itself,
returning to *To do*. People per column are yours to add by editing the rules. Spot these
templates by the "includes N automation rules" badge.
- Kanban templates get their own category chip ("Kanban") in the template gallery instead of
hiding under "quality".
- A project born from a kanban template opens as a board: its columns are laid out in one row
side by side rather than packed into two levels.
- The New project → From template dialog now shows template names, descriptions and
categories in the UI language (the English UI used to show Czech names).
- Export bundles the map's rules and import creates them again after remapping — a kanban
board survives deletion and being passed between instances. The import summary honestly
reports rules imported / skipped; whatever cannot be created on the new map is openly
skipped, nothing vanishes silently. Older exports without rules keep working unchanged.
- Privacy: the "without people" export strips everything personal from rules — assign-person
actions and notifications to a concrete e-mail are dropped, checklist assignees are emptied,
and a rule conditioned on a concrete person is left out entirely (roles like "node owner"
or positions stay). Import never lets an e-mail unknown to the target instance into the
rules.
- Docs: the rules page gained "Kanban from a template" and "Rules travel with the map"
sections (cs + en).
## v0.31 — 2026-08-15
**Kanban: a finished card moves to the next column by itself**
- Kanban move: a new rule action "move the node" (under a chosen goal, appended at the end of
its row — the rest of the map stays put, manual layout is preserved) and a new condition
"parent node" (catches cards under a specific column).
- The "Enable kanban" wizard (in the ⚡ overview and in the node's Automation category): pick
the column row, optionally a person per column, and it creates the whole chain of rules
"card under D1 marked Done → move it under D2, hand it to the column's person, return it to
To do". The generated rules are ordinary map rules — individually editable; enabling twice
on the same row is warned about.
- A moved card loses its "done" and returns to *To do*, so the next step always makes sense;
no loops are possible. Under the last column the card stays done and the case is closed.
- A map with active move rules shows a "Kanban" indicator in the toolbar instead of Arrange
(a board has nothing to rearrange; disabling the move rules brings Arrange back).
- "What changed" gains a **Moved** group (from → to, with column names) — card moves are
visible in the project history.
- Honest safeguards: moving to a vanished target, moving the apex, or a move that would
create a cycle is skipped with a plain reason in the run log — nothing crashes. Structural
rule actions are forbidden on the org-structure map.
- Node window, Automation category: reorganized into separate cards — rules on top, "Who
performs it" below, and the automation picker is a visible select from the agent registry
plus "other automation" as free text.
- API/MCP: the `move_node` action and the `parent` condition in `create_rule`/`update_rule`;
the docs rules page gained a Kanban section (cs + en).
> **Upgrade note.** A colleague with the map open sees someone else's card move after a short
> while via the "map has changed — reload" bar; structural changes are never silently merged
> into unsaved work.
## v0.30.1 — 2026-08-14
**Cards in the "by category" arrangement read in order**
- Templates with numbered steps (8D, FMEA…) rendered out of order in the "by category"
alignment style. The right column now reads top-down and the bottom row no longer slides
under the columns, so D1…D8 read in sequence.
- The whole "U" is centered on the true center of the apex circle; with a longer bottom row
the columns move outward, so layouts of 8+ cards get somewhat wider — the price of correct
ordering.
- Maps where the scrambled order was already saved ("baked in") are not fixed retroactively —
re-create them from the template.
## v0.30 — 2026-08-14
**Org structure, deputies and smarter rules**
- Organisation structure: your company as a tree of positions and appointed functions, with
holders and per-position deputies. Admins draw it, everyone can read it (user menu); the
deputies table in Organization settings does everything without the map — add, rename,
appoint, remove — and saves instantly.
- Deputies: a member's personal deputy as the fallback; position deputies from the org
structure take precedence. Removing a member vacates their positions automatically and
notifies the admins.
- Dynamic rule targets: "deputy of the responsible person", "holder of position X", "deputy
of position X" — resolved at run time. An unresolvable target is an honest skip in the run
log; the rule is never broken by it.
- Action target "On node": set status / assign / set deadline can aim at the trigger node,
its parent, or any specific node — a finished sub-step can start and staff the step above
it.
- "Overdue" now means *at least N days past*: it also catches deadlines that expired before
the rule existed, and fires once per deadline (a changed deadline may fire again).
- Node window: an Automation category (Behaviour merged in), the run log one click away from
the node panel, and rule changes from the same save appear instantly without a reload.
- Stress-free invitations: the set-your-password link is valid for 3 days instead of half an
hour, with a clear "request a new link" page when it expires.
- API/MCP: `GET /v1/org-structure` and the MCP `get_org_structure` tool (read-only); the docs
site gained an Automation category and an Org structure page (cs + en).
> **Upgrade note.** Existing "overdue" rules may catch up once on older deadlines after the
> upgrade.
## v0.29 — 2026-08-14
**Automation rules — "when X, do Y" — and a large node window**
- A built-in automation engine: rules of the form WHEN (6 triggers: node status · unblocked ·
deadline before/after · new node · attachment · scheduled time) → IF (AND conditions) → DO
(6 actions: set status / responsible person / deadline · create sub-nodes · send a
notification · run an agent). Rule runs are never counted or limited.
- Everything is available through the API and MCP too — an agent can create a rule by itself
(17 MCP tools).
- Rule templates: save a rule's shape once, load it in any map as a copy.
- A large node window with a left category menu for map editors, and a simplified window for
collaborators — no more one long scroll.
- Rule builder right in the map (the lightning button), a rules overview with a run log, and
badges on nodes; integrates with "Wait for children".
- Safeguards: chained rules stop at depth 3, caps on runs and saved rules, a
`KB_RULES_DISABLED` kill switch, and a broken rule e-mails the map owner once.
## v0.28.1 — 2026-08-14
**The required goal picker says so**
- The task dialog's goal field now shows the "Select a goal" prompt — the required field used
to just look empty.
## v0.28 — 2026-08-14
**A task always lives on a concrete goal**
- A task must belong to a concrete goal in the map. The project apex does not accept tasks —
it completes by its goals completing. The task dialog has a required goal picker (the apex
is not offered), and "detach from goal" was replaced by moving to another goal.
- Importing maps: tasks from backups without a valid goal are skipped and reported
(`tasks_skipped`); the "map is a tree" guarantee now also holds for imports with positions
and detached cycles.
> **Upgrade notes.** API change: v1 `POST /tasks` and MCP `add_task` now **require** a
> `node_id` of an existing non-apex goal (otherwise 400 with a clear message). A migration
> moves existing apex/goalless tasks into a new "Unsorted tasks" goal (titled in the owner's
> language; if everything is done the goal is created already completed, so project progress
> never drops). Nothing is deleted.
## v0.27 — 2026-08-13
**Map conflicts without losing work, and Midnight as the default skin**
- The conflict dialog (409) no longer throws away work in progress: it offers "Keep my
changes" (a conscious takeover on a fresh base) and "Download a backup (JSON)" alongside
loading the current version.
- Background change watch: the editor cheaply asks for just the map's version every 45
seconds and shows a gentle bar when someone else changed the map — before you start
typing. Status-only changes still merge silently.
- The instance default skin is Midnight (dark in both modes).
## v0.26 — 2026-08-13
**Map readability: three font sizes on one button**
- A new "Readability" button in the map toolbar cycles three node font sizes the same way
Arrange does: normal → larger → name only. The default is now "larger" — anyone who never
chose anything sees the map more legibly right away; saved choices are untouched.
- "Name only": a large name (24 px over three lines), the description hides behind a "…"
marker with the full text in a tooltip, the progress bar and apex badges disappear — only
what things are called remains.
- The choice is per device: keep normal on the monitor, switch the phone to large. It also
works in read-only maps — enlarging text must not depend on edit rights.
- Long names get a tooltip with the full text (the bigger the font, the less fits the card).
- Autosave stopped sending saves that change nothing. Previously a mere click on a node
saved the map: it jumped in the "recently edited" ordering and desynced a colleague's
version in a shared map (409 conflicts).
## v0.25.1 — 2026-08-13
**A map is a tree: cycles and second parents can no longer be created**
- A connection that would create a cycle, or give a goal a second parent, can no longer be
made — and the app says why, instead of a silent nothing that looks like a malfunction.
(Such maps used to freeze the browser tab at 100 % CPU on open.)
- A damaged map announces itself on open and offers a Repair button. It detaches only the
extra edges, keeps every goal, and can be undone.
- A newly drawn edge can now be undone — the Undo button used to ignore it.
- The server rejects a cyclic map through the API as well — but only *new* damage: a map
already damaged today still saves, so its owner is never locked out of it.
- New script `product/audit-strom-map.js` lists how many maps in the instance are damaged.
It only reads, never changes anything.
## v0.25 — 2026-08-13
**Arrange finally does something: compact styles, cards around the center, and a lock**
- On some map shapes the Arrange button did literally nothing — all three styles returned
bit-for-bit identical positions. Fixed exactly where it matters most: fresh maps and deep
ones.
- A row of cards without sub-goals wraps into two levels, the lower cards sitting in the gaps
of the upper ones. Six cards: 1500 → 675 px, half the width.
- New "around the center" style: on a fresh map the cards walk around the project from the
left, below and right instead of one wide row. On a deep map, categories split into two
bands as before.
- Tighter card spacing (80 → 50 px) and smaller gaps between wrapped rows; connector bends
near the apex now line up instead of each breaking somewhere else.
- After arranging, the view centers itself — and Arrange can be undone.
- Style lock: hold the button to lock the style for all your maps; the button changes colour.
The lock is stored on your account, so it applies on mobile too. It never touches saved
maps — it only redraws — and it does not apply in someone else's shared map, so it cannot
overwrite the owner's layout.
- Arrange is now also available in "My map", the chosen style survives switching between
portrait and landscape, sticks to the specific map (not the browser), and AI rearranging
respects it. A new map is born in the locked style, otherwise compact.
- Fixed along the way: a map with a cycle in its edges froze the tab at 100 % CPU (a step cap
now prevents the freeze), cards taller than 240 px overlapped in wrapped rows, and a node
named `__proto__` crashed the layout computation.
## v0.24 — 2026-08-13
**See and find: card icons in every skin, and the user menu in the map**
- The card header icons (time tracking, detach, stash, delete) are visible in **all** skins.
In "Midnight" they vanished completely — the header had a hard-coded colour, leaving light
icons on a light strip. The header now follows the chosen skin, and the "Ruby" skin was
lightened so its icons are readable too.
- A light/dark toggle directly in the skin picker — a skin is chosen for both modes, and now
you can see both without walking into a map.
- The user menu is available in the map as well. The map was the only screen without a
header, so account, skin and language were missing there and the "top right menu" guidance
did not apply. The menu is one shared component now, so the two places cannot drift apart;
the ⋮ button keeps the map actions.
- The onboarding task "Change your skin" mentions both places (the menu under your name and
the palette at the bottom left of the map).
- Fixed a dead "back to the simplified view" button in the mobile header.
## v0.23 — 2026-08-11
**External contacts, three arrange styles, new map toolbars**
- External contacts: a directory of people outside the system (accountants, suppliers) who
can be assigned goals and tasks with deadlines. They never receive anything — it is purely
internal tracking; overdue items are announced to the assigner in the daily summary.
Contacts are visible org-wide, with optional private contacts (anonymous on shared maps).
They are created right from the responsible-person picker; the name lives only in the
directory, so deleting a contact leaves data readable.
- Arrange cycles three styles on one button: wide (classic) → compact (alternating levels,
about 30 % narrower) → by category (second-row nodes split the project into two bands).
The button always shows the style currently applied to the map.
- Map toolbars reworked on desktop and mobile: search, the My-tasks filter and the Dashboard
moved to the left rail under the stash and timer; the top bar has view controls on the left
(direction · arrange · fit-to-map) and creation and messages on the right.
- "My map" mirrors the project structure: a branch per project, your goals hanging on the
map's real intermediate nodes — no more one wide fan.
- A fresh instance greets you with administrator sign-up instead of a "Welcome back" login
into the void.
- My day: expanded by default on Projects, collapsed on Tasks, each page remembering its own
choice.
- Link attachments recognise Gmail (envelope icon) and Google Drive/Docs (triangle) and show
a readable name instead of a truncated URL.
- Documentation: a new External contacts page, sections on arrange styles, lite in the skins
guide, and up-to-date toolbar descriptions — cs and en, with regenerated screenshots.
## v0.22 — 2026-08-11
**Without mail, passwords are reset by the admin — not by a promise**
- Without SMTP configured, "Forgot your password?" is no longer offered (neither the link nor
the page) — the server used to answer "all right" while no message could ever arrive, so a
forgotten password meant a lost account. Instead, the app points to the instance
administrator. With SMTP configured nothing changes.
- An admin can reset a team member's password: without mail they get a temporary password to
hand over, with mail a regular reset link is sent. Deliberately **not** for themselves and
not for another admin — otherwise two admins could take the instance over from each other.
- The affected person always gets a "someone changed your password" alert that cannot be
turned off — it is a defence against a silent account takeover, not a routine notification.
- Documentation (cs + en): what changes without SMTP, the forgotten-password procedure
including for the admin (on self-host via the PocketBase console), and an AI model
recommendation based on measurements — gpt-oss:20b verified, with hardware needs and
response times.
## v0.21.1 — 2026-08-08
**Fixes from real-world use: invitations, returned work, planning**
- The invitation e-mail now says where it is inviting you: the subject and heading name the
organisation, the text spells out the word you enter on the sign-in gateway, and it carries
a permanent sign-in address for your bookmarks. Below the button there is what to do once
the one-time link expires ("Forgotten password"). The logo and footer link lead to your own
instance.
- Work that was sent back no longer also shows as finished — undoing a completion used to
leave the item both back among the tasks and still in "Done today".
- Your plan ("when I want to deal with it") now decides which section an item lands in, even
for due dates today and tomorrow. The due date itself never changes and stays visible; a
plan never pushes an item more than a week ahead, so an approaching deadline cannot hide.
- Daily summaries and deadline notices respect the plan: the summary does not scold you about
work you consciously postponed, and "today/tomorrow" notices stay quiet until the deadline
actually passes. An overdue deadline is still reported — delays must not hide.
## v0.21 — 2026-08-08
**A tidier user menu and real names**
- The user menu is organised into sections; Clients moved under Time tracking.
- New "My account" page: name, display name/nickname, and password change.
- Nodes and sharing show people's names instead of e-mail addresses.
## v0.20.1 — 2026-08-08
**Trial countdown for admins, AI that apologises, invitations that say who invited you**
- Organisation admins see the trial-days countdown for the whole trial — a subtle bar above
the header. Members still get the prominent notice only in the last week, plus an
explanation after expiry.
- AI map generation no longer waits forever. On an outage the client gives up after 90
seconds with an apology instead of hanging. Audio transcription keeps its own, more
generous limit — and a server-side cap that used to cut long transcriptions short was fixed.
- AI outages are recorded in an error journal (who, which feature, why, when) so the operator
can evaluate reliability. **Map content is never stored there.**
- The invitation e-mail now says who invited you, by name and address.
## v0.20 — 2026-08-08
**A due date is an agreement**
- An existing due date on a goal or task can only be changed by the person who set it, or by
the project owner. Setting the *first* due date stays free. Enforced on every write path —
app, REST and the v1/MCP API.
- A goal carrying an assigned task can only be removed by its assigner or the owner. That
covers delete, the Delete key, stashing and conversion to a note.
- New sharing level **Collaborate**: the colleague sees the map, completes only their own
tasks, comments — and changes nothing else.
- **Due date change requests.** If you may not change a date, propose a new one with a reason.
The assigner is notified and approves simply by setting the date, or declines. The requester
always learns the outcome.
- Deleting a task now leaves a trace in "What changed"; delete and stash controls only appear
for people the server would actually allow.
- Fixes: editors on maps with 2+ shares can create tasks again, and so can team members with
edit access.
> **Upgrade note.** Assigning a task to someone who has no access to the map now shares it at
> the **Collaborate** level. Previously that silently granted full edit rights. Existing shares
> are not changed — this only affects assignments made from now on.
## v0.19.1 — 2026-08-07
- **An automation naming an unregistered agent now stays quiet.** If a goal names an
automation that matches no registered, enabled agent, the step simply does not run — no more
confusing "Agent not found" notification. It is read as a note that a machine does this step,
not as an instruction. A *registered but disabled* agent still reports the failure.
- Documentation: a new section lists plainly when an automation runs.
- Clicking the header logo takes you home from anywhere.
## v0.19 — 2026-08-07
- One unified AI dialog for the whole project, instead of separate entry points.
- The AI answers in the user's language.
> **Upgrade note.** On mobile, the simplified (lite) view has no AI — that is deliberate, not
> a regression.
## v0.18 — 2026-08-07
**Everything that an HTTPS domain unlocks**
- **Sign in with Google.** Enabled by configuring `KB_GOOGLE_CLIENT_ID` / `KB_GOOGLE_CLIENT_SECRET`
— without them the button never appears. Works on instances with an activation code too, and
the seat limit is still enforced.
- **killBottleneck over MCP, remotely.** Every instance exposes MCP at `/mcp` (Streamable HTTP),
so Claude Code and Claude Desktop connect with nothing installed locally — just an API key.
Same nine tools, same limits and authorisation as the local server.
- **claude.ai connectors.** The instance speaks OAuth (client registration, PKCE). The issued
token appears under "API keys" in the app, where you can revoke it.
- **Google Drive attachments.** A step can offer "Pick from Drive" — the chosen file is added
as a **link**, nothing is uploaded and the file stays on your Drive. Shown only when a Picker
API key is configured.
> **For operators.** Google login needs the redirect URI `https://DOMAIN/api/oauth2-redirect`
> in the Google Cloud Console. The Drive picker additionally needs the domain under
> "JavaScript origins" and the Picker API enabled with a referrer-restricted key.
## v0.17.1 — 2026-08-07
- Fixes found in a live click-test: invitations, the idea stash, mobile, and a bug where an
AI-generated map could be saved without its nodes.
## v0.17 — 2026-08-06
- A release focused on our hosted cloud (sign-up and plans); nothing changes for self-host.
## v0.16.1 — 2026-08-05
- Fixes from the second review round.
## v0.16 — 2026-08-05
- Lighter mobile view, notification budgets, AI available in the cloud.
---
## Earlier releases
| Version | Date | Headline |
| --- | --- | --- |
| v0.15 | 2026-08-02 | Protected apex, subgoals without collisions, templates in CZ + EN |
| v0.14 | 2026-07-31 | The most important task of the day |
| v0.13.2 | 2026-07-31 | Round apex, straight edges, import from Asana/Trello |
| v0.13.1 | 2026-07-31 | Export in the current skin, bilingual skins |
| v0.13 | 2026-07-31 | Visual skins |
| v0.12 | 2026-07-29 | English, licence, and version checking — ready to go public |
| v0.11 | 2026-07-28 | Renamed to killBottleneck; Cloud Lite; attachments as links |
| v0.10 | 2026-07-27 | Two faces of one app: the map for those who steer, a list for those who do |
| v0.9 | 2026-07-26 | Automations in the map: who performs a step, attachments as a trigger |
| v0.8 | 2026-07-25 | v1 API and the MCP server — AI can build maps |
| v0.7 | 2026-07-24 | Tasks always live in a project; My map (To do / Assigned by me) |
| v0.6 | 2026-07-24 | Bilingual EN + CZ |
| v0.5 | 2026-07-23 | Responsive map direction on mobile |
| v0.4 | 2026-07-22 | My day, time tracking, exports with sharing |
| v0.3 | 2026-07-20 | Numbered series from templates, archive, unified header |
| v0.2 | 2026-07-19 | Project and node colours, emoji icons |
| v0.1 | 2026-07-10 | First versioned state |