v0.6.0

Added

• Writing and proofreading, in English and Bulgarian. A new Writing page,
  and the same in the chat ("proofread this", "make it more formal"). The
  proofreading uses rules rather than a model — instant, offline, nothing
  leaves the computer — and every correction says why: doubled words,
  spacing, Latin letters hidden in Cyrillic words, в/във and с/със, ѝ, „не“
  apart from a verb, по-/най-, the full article after a preposition, the comma
  before че, защото, но and който, „…“ quotes; a/an, could of, more then, I.
  Sure corrections are ticked and made in one press; the rest are yours to
  decide. Choose a tone (formal, friendly, concise, persuasive, simple,
  polite) and a word limit to have them checked too, or have the text
  rewritten in that tone and see the changes before keeping them. Compare
  two versions word by word, on the page or with "what changed?".

• Translation with your own glossary. Translate a text, or a file into a
  new file beside it with its layout kept: Word and OpenDocument (styles,
  tables, headers, pictures), subtitles (timings kept), Markdown (code left
  alone) and plain text; the original is never replaced. Keep a glossary of
  agreed translations — add terms one by one, paste them from a spreadsheet,
  or save them as one — and every translation follows it and is checked: a
  term not translated as agreed, or a number, address or link that went
  missing, is listed with the result. A new Translation page, in English and
  Bulgarian, and a translator specialist for longer jobs.

• Scans, photos and spreadsheets. Scans and photos — a scanned contract,
  a photo of a receipt, a PDF with no text in it — are read on this computer,
  by Tesseract or by a local model that can see, and say they came from a
  scan. Spreadsheets are written as real .xlsx files (bold frozen header,
  numbers as numbers) and read back as rows with their numbers, every sheet
  by name; OpenDocument files are read without LibreOffice. Tesseract and the
  PDF tools can be installed from the list of what a task needs.

• Mail, sorted, with replies that wait for your yes. A new Mail page sorts
  new mail from its headers — suspicious first, then what needs a reply, then
  people, then newsletters and notifications by sender — from your mail
  server or from Thunderbird. Replies are drafted in the message's language
  and tone and kept waiting; each is sent only when you press Send on it, as
  it stands, and is recorded in What I did. Sending a draft from the chat
  always asks, even when everything else is allowed. Optionally, a one-line
  summary of new mail every morning.

• Plan: what is next, to-do lists and your online calendar. A new Plan
  page puts the diary, your online calendar, reminders and due to-dos in one
  time order, overdue first. Keep to-do lists by name, with an optional due
  day, ticked off in one click. Connect a CalDAV calendar (Nextcloud,
  Fastmail, iCloud and others) with an app password: the server is checked
  before anything is kept, its appointments appear under What is next, and
  new ones can be added — never changed or deleted. Also in the chat: "what
  is next?", "put milk on the shopping list".

• Answers say where they came from. Each answer lists its sources — things
  it remembers with the file they came from, files and web pages it read,
  search results — numbered, and a written answer ends each sentence that
  relies on one with its number, like [2]. A number that points nowhere is
  taken out before the answer is shown. Open a web source in your browser or
  a file's folder from the list; the list stays with the conversation.

• What I did and why. A new page lists every action, every action put to
  you and every refusal, with the request behind it, the specialist, the
  model and why it was chosen, what it was given and returned, and whether
  anything left the computer. Filter by period, by changes, or by what left.

• Control. The permissions page is now Control: every capability says
  whether it can send anything off this computer and what, has an on/off
  switch, and can be filtered; citing sources is switched here too. These
  pages, and the sources, are in English and Bulgarian, following the
  language you speak to it in.

• Each step gets the model its work needs. A job's steps are thought with
  the model their expert's work calls for — code, hard thinking, speed,
  pictures — chosen from what is installed at that moment, with the reason
  shown; a model that keeps failing gives way to the next.

• One question at a time. When it needs to ask, it asks one question with
  two to four answers to click, the recommended one first; reply with a
  number, a word or your own sentence.

• Starting knowledge inside the program. A new brain gets short guides in
  its knowledge folder (getting started, portable copy and USB, the scam
  guard, experts and models, trusted sources). Updates refresh the guides you
  have not changed and never touch the ones you have.

Changed

• Specialists for every field, and only when needed. Business strategist,
  finance analyst, accountant, legal specialist, marketing specialist and
  music producer join the developers, writer, designer and researcher, each
  with only the tools its work needs.
• Original and meaningful before finished. Whatever a job produces — a
  game, a logo, a jingle, a story, a strategy, a contract draft — is checked
  by the specialists of the fields it touches (trademarks, look, sound, game
  rules, text and meaning, money, tax, law, business), side by side, and only
  for the sides it has; why each check was or was not used is in the task,
  and the findings have their own "Originality and meaning" section in the
  final report. New guides: "How I work" and
  "Originality and licences", in English and Bulgarian.
• Works on goals on its own. One setting, agreed to once: a goal is
  planned, staffed, provisioned and finished end to end. When no model here
  fits a step, the largest one that fits the computer's memory and disk is
  downloaded and its checksum verified, falling back to the next on failure;
  user-level tools are installed from their official sources with their
  published checksums; each goal that makes something works in a folder of
  its own; the report lists what was put on the computer. Only four things
  are still asked: a payment or a paid service, sending or publishing in your
  name, deleting files outside the goal's folder, and a change to the whole
  system. Only what it downloaded itself is ever offered for removal.
• The goal-mode consent says what privacy does not cover. Models, tools
  and integrations are downloaded from the internet even when privacy is
  "nothing leaves this machine"; the consent says so, and that those
  downloads send nothing about you. Settings show the same words the program
  enforces.
• Integrations a goal needs are found and added on their own. From the
  built-in catalogue first, then the official MCP registry — only from
  trusted publishers (add your own in integrations/trusted.json), pinned to
  one version from npm or PyPI, running on this computer and needing no
  account or secret, and only while privacy allows it. Each is reviewed
  before anything is downloaded, given to the one specialist that needs it,
  and the tools the server says destroy are not offered. The registry is
  searched with a short word for the kind of tool, never your words.
• Unused downloads are removed on their own. Once a day, models, tools
  and integrations that goal mode downloaded and nobody has used for 30 days
  are removed and the removal is proven; never anything you installed or set
  by name, and nothing while a job is under way. A failed removal is tried
  again the next day.
• A goal is tested from the request to the report. "Make me a space
  strategy game", with goal mode on and privacy set to this computer only,
  runs on a brain of its own against a fake model library and tool source:
  the plan, the specialists, a model and a tool downloaded and checked, the
  work in the goal's own folder, the originality checks and one report —
  with nobody asked anything.

Fixed

• The diary said appointment times in UTC rather than local time, so an
  appointment at two in Sofia was read as noon and a whole day as the evening
  before.

• On a computer without graphics acceleration goal mode no longer downloads
  the largest model that fits in memory (minutes per answer); it chooses one
  that answers in reasonable time.

• When nothing here could write code, goal mode now downloads a model that
  can, instead of stopping the goal.

• The general assistant is no longer limited to a personal assistant's mail
  and diary tools, and a specialist is offered only the tools it may use, so
  steps no longer go round between specialists until they fail.

• A "strategy game" no longer brings a business strategist; music a plan
  makes brings the sound check even when the request never said "sound".

• Installs in goal mode wait for the result, are recorded and can be removed
  again; the report says work was done without asking because of goal mode.

• A data race when steps of one job ran side by side.

• Originality checks with real tools. Names are searched in trademark
  registers, app and game stores and the web; a game's one sentence is
  compared with the closest titles; pictures are compared with reference
  pictures you give (resized, recoloured and mirrored copies are found);
  melodies in MIDI files and code are compared, in any key, with well-known
  public-domain themes. Findings go into a report — item, aspect, finding,
  risk, action — saved as a table and a CSV sheet to download. "Audit the
  originality of " checks a whole catalogue, each item only for what
  it has, side by side; your own works are reported, never changed.

• Fixed by itself. When a check finds something in what the assistant
  made, the one who made it changes it and the same specialist checks again;
  the report says what was changed.

• Plain wording on the screens. The brain is now explained where its
  copies are kept: it is the assistant's knowledge base, everything it has
  learned, remembered and been told, in one folder. Team members are called
  specialists, learned ways of working are called procedures, and "ships
  with" became "included with".

• A moved brain does not learn its own knowledge again. Files inside the
  brain's folder are remembered relative to it (@brain/knowledge/…), and
  what was learned before is rewritten that way on the next start, so a brain
  on a USB stick or another computer keeps one copy of each fact.

• Each expert runs only its own tools, also when everything is allowed.
  A step that reaches for another expert's tool is handed, without asking
  you, to an expert who has it (a specialist first, the general assistant
  only when no specialist does).

• The built-in playbooks are only what every user needs: asking, planning a
  goal, mail, documents and the scam guard. Procedures specific to one
  business are no longer shipped in the program.

Full Changelog: v0.5.0...v0.6.0