Kriko 0.2.6

Kriko 0.2.6, built from source commit 9279af0.

The Windows x86_64 desktop build. The installer installs for the current user
and needs no administrator prompt; the portable ZIP needs no install at all.
Installed catalogs update themselves weekly.

What changed

• Fixes from a bug hunt over the whole app: a price with two numbers in it
  ("from 10 to 30", "-50% 199,90 TL") is no longer read as one glued price, and
  a budget or review count written the way a shop prints it ("15.000", "1,234")
  is read as the number it is. A watched page keeps its address as written.
  Forgetting purchases or watches stays forgotten on a paired device. A cart's
  lines are counted once, the price beside a product's name is the one it reads,
  and a page that changes without loading is read again.
• Safer fetching: a page that redirects to this computer or your own network is
  refused, a page built to stall the reader is bounded, and one page's broken
  data fails alone.
• Relay and sync: a large body is refused before it is read, and one bad item no
  longer stops the rest of a sync.
• The app: For you keeps what you type, a second restore keeps the state just
  before it, saved checks are found by words with ü and ı, a failed settings
  save keeps its message, Start Ollama says why it could not, and the queue and
  Compare no longer call a job failed because one read of it was slow.

Catalogs published beside the installer: cars.kpack, drill.kpack, and
packs.json, the index an installed Kriko polls.

These desktop builds are unsigned and not notarized. Windows SmartScreen may
warn before running the installer (More info → Run anyway).

Kriko keeps everything in ~/.kriko: knowledge.sqlite for installed
catalogs, app.sqlite for your own history.

Kriko 0.2.5

• An update now says what version it is. After 0.2.4 was installed over
  0.2.3, Kriko kept reporting 0.2.3 and kept offering 0.2.4 to itself, so Update
  now reinstalled the same thing. The installer laid the new program over the
  old one and left the old one's version record beside it, and Kriko read the
  wrong one. Kriko now reads the newest version record it finds, and the
  installer clears the engine's own folder before it installs.
• Uninstall removes the engine's folder. The same leftover files were not on
  the uninstaller's list, so the folder stayed behind and the uninstall could
  fail. The uninstaller now clears that folder too, after it stops Kriko. Your
  catalogs and chats are not in it: they stay in .kriko in your profile.

If you are on 0.2.4 and it still shows 0.2.3 in Settings, About, quit Kriko from
the tray and run the 0.2.5 installer once by hand. That install removes the old
record.

Kriko 0.2.4

• Kriko starts agents when you need them. Opening the app no longer
  launches your coding agents or asks them for their model lists. The window
  reads a short list of the agents you have, and the full discovery happens when
  you open Settings. An agent that sits idle is now closed after 90 seconds
  (20 on a machine with little memory), and Kriko keeps at most one warm. Closing
  an idle agent no longer holds up the one you are using. The agent picker
  refreshes when you change a setting or a key, without resetting a conversation.
• A local model no longer shows its thinking or a raw search request. Some
  local models printed their reasoning, or answered a request to search with a
  block of structured text. Kriko now hides the reasoning, reads the common
  shapes of a search request and runs the search. If a model's request cannot be
  read, Kriko searches for your own words.
• Use any OpenAI-compatible API as an agent. Add your key in Keys, and the
  endpoint and the model ID in Settings, Research. The agent can use the Exa or
  Tavily search you chose.
  Your key is never shown again, and a reply that cannot be read is refused
  instead of guessed.
• Choose the market and the language. Settings, Research, has a language for
  answers, a language for searches and a market (a country or the European
  Union). Chat, quick looks, research and follow-up questions all use them. The
  agent is also told to match the length of the answer to what is at stake, to
  require exact fit for a part (a PC part or a vehicle part), and to report what a
  forum alleges as an allegation, not as seller evidence.
• See what Kriko searched and read. A run lists the exact addresses it
  searched, how many pages it tried and how many it read. These are three
  numbers, not one. A cancelled read is not counted as read. The log of a run
  stays with the conversation when you open it again.
• A finished queue clears, and keeps its results. When every research in a
  batch is done, the queue empties. The saved checks stay in your history, and
  anything waiting, paused, failed or added later stays in the queue.
• Chats get a title, and suggest what to ask next. The agent names a new
  conversation and offers up to three related questions. A click puts one in the
  message box, where you can edit it before you send. Your answer is saved first,
  and a title you chose yourself is kept.
• Kriko files a product under its catalog with your agent. When a quick look
  is saved, Kriko matches it to an installed catalog by what it already knows. If
  that is not enough, the agent you picked reads the saved facts and the catalogs
  and proposes one, with the exact words it relied on. Kriko accepts it only when
  the catalog is installed, the confidence is high enough and the words are on the
  page. Otherwise the product stays uncategorized, and you can retry or choose.
  History keeps the method, the evidence and the catalog's version. An agent that
  works through MCP can list the catalogs and submit its own decision.
• Optional: Jev from TypeSafe can do the filing instead. Off by default. You
  turn it on in Settings, Research, and add a TypeSafe key in Keys. When Kriko
  cannot decide by itself, it sends the product title, the saved facts and the
  list of your installed catalogs to TypeSafe. It does not search the web. It is a
  paid service, and Kriko records the usage the service reports.

Also in this release: a benchmark for part compatibility (python -m app.compatbench),
which grades a model on five invented cases and checks every passage it cites.
It measures nothing until you run it against a model.

Kriko 0.2.3

• Kriko opens in a few seconds, not fifteen. The engine was one file that
  unpacked itself into a temporary folder on every start, and on Windows that
  alone took 10 to 13 seconds. It is now a program with its files beside it,
  unpacked once by the installer. The shell also no longer waits for the search
  service before it starts the engine.

• The window says it is starting. Until the engine answers, the window shows
  the Kriko mark and "Starting Kriko…", not an empty page.

• Kriko updates itself. A few seconds after it opens, and every six hours
  after that, Kriko looks for a newer release. When there is one, a note offers
  Update now or Later. Update now downloads the installer, checks it against the
  checksums published with the release, closes Kriko, installs and opens it
  again. Your chats and catalogs stay where they are. Settings, About, shows the
  state and has a Check now button. A copy that was not set up by the installer
  (the portable zip) is pointed to the download page instead.

  This starts with 0.2.3: to get from 0.2.2 or earlier to this version, run the
  installer once by hand.

• Chat finds the exact product and checks the seller. A request such as
  "can you find these for me?" now searches the web. Before this change, a
  request in that form received an answer from memory, with no source. The
  answer now follows these rules:

  • Kriko reads the pages of popular and credible sites first. The list is the
    file sellers.json. Add your own sites in ~/.kriko/sellers.json.
  • A code (a part number, a model number, an oil grade) is confirmed only when
    two trusted sites print it. The answer says so when it is not.
  • A marketplace is only as good as its seller. The answer names the seller and
    the score that the page shows, or it says the seller is unknown.
  • A page that calls itself official makes a claim. The answer reports it as
    the claim of the page.
  • The answer uses the language of your question.

• The chat follows new text. While an answer streams, the thread stays at
  the bottom. If you scroll up to read, it leaves you where you are.

• Researches in the queue run side by side. Every research waited on one
  shared draft claim, so the queue ran one at a time whatever the setting said.
  Several now run at once. Two researches on the same catalog still take turns,
  on purpose.

• Learn this page works, and offers itself. The panel offers Learn on a page
  of a shop Kriko does not know, including a real product page on an unknown
  shop. Once for each site, the browser extension suggests it when a page looks
  learnable.

• Compare leads with the problems. Each product shows its risks first, and
  its specifications in a table you can open.

• A quick look is not its own source. Specifications that only repeat the
  listing you gave are dropped, and the agent is told to check other shops.

• "Original" in a title is not a check. When you ask whether a seller or a
  part is genuine, the agent looks at the seller and at the maker's authorised
  dealers, and no longer counts the word in the product's title.

Also in this release, outside the installer: the website has been redesigned
around how Kriko finds the concern before you buy, and kriko-mobile/ holds the
Kriko reader as an Expo app (run it in Expo Go).

Kriko 0.2.2

• Each answer shows which agent wrote it. Claude Code, Codex, Antigravity,
  Mistral and the local model each have a mark of their own, above the answer
  and in the agent pick. An answer from an earlier version has no mark; Kriko
  never recorded who wrote it.
• Copy, Retry, Edit. Under an answer: Copy takes its words, and on the
  newest answer Retry asks the same question again. The newest question has an
  Edit button that puts it back in the box. Up in an empty box brings back what
  you asked before. A Latest button shows when you have scrolled away from the
  bottom of a long answer.
• Rename, pin, search and copy a chat. Each chat in the list has a
  "…" button. Rename (or double-click the name), Pin to keep it at the top, Copy
  as Markdown, Delete. The box above the list finds a chat by its name or by
  anything said in it.
• Links in an answer open in your browser. In the desktop window a link did
  nothing. Only web addresses (http and https) are opened.
• No sample chats while Kriko starts. For the seconds the app looks for its
  engine, the home screen and the lists showed demo rows, then swapped them for
  yours. They now wait.
• The folded sidebar lines up. The + for a new chat sat to the left of the
  other icons; the logo, the fold button, the + and the four pages share one
  centre line.

Kriko 0.2.1

• Learn this page. The research card in the extension panel has a button
  that goes through the page you are on and learns what to take from it. It
  shows what it took and what each thing was taken as. The model on this
  computer answers at once. Another agent runs as a job the panel follows.
• Pick the agent in the panel. The research card lists the agents that
  are ready, starts on the one the app uses, and remembers your pick. It
  names the cost before you press anything.
• A research search starts from the product, not the page title. A title
  such as "Samsung Galaxy A54 256 GB Beyaz Fiyatları, Özellikleri ve
  Yorumları | En Ucuzu Akakçe" was searched whole, so the searches found the
  shop and its ranking words instead of the phone. The shop name and the
  words a shop adds to rank are taken off before any search is made.
• A page about your phone is read, even if it never names the colour. A
  review of the Galaxy A54 was thrown away because it did not repeat the 256 GB
  and the colour a listing's title adds. A model code such as A54 with another
  word of the name is enough to read the page. Each finding still has to show
  it applies.
• The model on this computer maps a page correctly. It took a pack's name
  for a key, and what it wrote was refused. It is given the key names alone.
• Agents no longer fail with "does not declare per-run model selection".
  Kriko asks each agent's program which flags it has. One slow first answer
  was remembered until you closed the app, so Claude Code and Antigravity were
  refused for flags they have. An empty answer is now asked again after a short
  pause.
• Chat no longer ends on "SEARCH: …". When an agent asked for a search in
  more than 14 words, the request was shown as the answer and nothing was
  searched. It is shortened and the search runs.

Kriko 0.2.0

• The desktop app is new. It is built on Tauri now, with a blue look, a
  dither sky, and a dark and a light theme. Motion has three levels: Calm,
  Standard and Vivid.
• Chat keeps your chats. Press New chat to start one. Recent chats are in
  the sidebar. Deleting a chat asks first.
• A tour on the first start. It shows the box where you ask, the agent
  pick, where Kriko looks, and the places in the sidebar. Take it again from
  Settings.
• The extension panel has the same look and its own tour. It shows once,
  on the first product page. The options page can show it again.
• Local models. The Local model panel shows your graphics memory and the
  models Ollama holds. You can browse Hugging Face for GGUF models with
  filters for size, licence and sort order. A filter can keep only the models
  that fit your graphics card. A download starts only when you press Get.
• vLLM on CUDA 13. Kriko reads the CUDA version of your driver and shows
  the steps for the right build. It does not install anything. On Windows the
  steps use WSL 2. Kriko finds the server when it runs.
• Catalogs and Sites panels. Remove asks before it removes.
• The sidebar, Compare and Queue use the new layout. The installer is
  built from kriko-ui/package.ps1.

Unsigned. Windows SmartScreen warns on first run: More info, then Run anyway.

The installer is kriko-0.2.0-x86_64-setup.exe. It needs no administrator rights.

Kriko 0.1.2

Changelog

Reader-visible changes for releases older than the three shown in the
README. Newest first. Published versions are immutable; a
correction is a newer version (see the update protocol).

0.1.2

• Chat answers are formatted. An answer from an agent is markdown. A bold
  term is bold. A bullet list is a list. A table is a table. A link opens the
  page it names. Before this release the screen drew the characters: every
  asterisk, pipe and bracket reached the reader as itself.
• You can select and copy an answer. Drag over the words. Press Ctrl+C. The
  selection crosses a bullet and the paragraph under it, because the
  transcript holds one selection rather than one for each block.
• One status line while an agent works. The running turn showed a moving
  mark, a phase word and a second status line at the same time. The mark and
  one line remain.
• A product card under an answer. When a turn reads your catalogs, the
  answer now carries a card for each product it found. The card shows the name
  the catalog uses, the identity it is recorded under, and the number of
  recorded risks. Open reads the product in Browse. Add to Compare puts
  it beside the others and saves its check. Read again reads the product
  now, because an answer is a reading of the catalog at the moment it was
  written.
• A calmer sidebar. The rows have one rhythm. The fold control moved to the
  brand row, where it reads as this rail's control. A real edge separates the
  rail from the page; a hairline disappeared on a bright screen.
• The model store is a list. One full-width row for each model, with one
  primary action at the right edge. The name is the largest thing in the row.
  A size picker appears only on a model that is not installed and offers more
  than one size.
• The local model lives with the agents. Agents now shows a marketplace
  of every way to answer a question: the coding agents on this machine, the
  local model, and the paid APIs you hold a key for. Each row says what it is,
  where Kriko found it, and its state. It offers one action. The local model
  panel shows the model, the server and the number of models on the machine.
  Local LLM left the rail and keeps its own screen, which Agents opens.
• The light theme is correct. The hero band kept its dark image but its
  background followed the theme, so on the white ground it became a pale smear
  with light words on it. The band keeps its own colours in both themes. A new
  EDGE token separates two surfaces, where a hairline is only right for a
  rule inside one surface.
• The browser extension is ready for the Web Store. Version 0.4.2. Escape
  now closes the panel. One loading indicator replaces two spinners and a
  skeleton. The endless sky movement is one pass. A pack's tone value is
  escaped. The result cache is keyed by a digest, so the reader's own listing
  addresses are no longer written to disk as keys. The options page keeps a
  visible focus ring, and it no longer claims that nothing leaves the computer.
• The agents are faster and they need less of the internet. The plane
  decision took 1114 ms on a cold start and now takes 264 ms; the second and
  third question in one turn take 1.8 ms. One request on the loopback address
  took 17.19 ms and now takes 1.43 ms. A read of the installed catalogs used
  159 statements and now uses 6. Three of the four coding agents declare no
  page reader, so no internet server can attach to them. The rule that a paid
  API runs only after you pick it is unchanged.
• A follow-up question carries less and knows more. The brief for a saved
  check fell from 37,229 characters to 8,752. It still carries the product, how
  far it has been used, and every recorded risk with its page.
• Kriko can work out how to read a new site. When no installed catalog can
  read a page, the local model is asked to map the page's own labels. The
  result is checked before it is stored, and it is cached by the page's shape
  rather than by its address, so one inference serves every listing on that
  site. Nothing is recorded when no model is ready, and the reader can turn an
  inferred site off.
• Known limits: the desktop screens were compiled, unit-tested and observed in
  a window on Windows only. The agent speed figures come from one machine and
  exclude the model's own inference time. Nothing was checked on macOS. The
  browser extension is not yet published to the Web Store.

Kriko 0.1.1

Kriko 0.1.1

Windows only for this release: the MSI and the portable ZIP.

What is new

• The desktop interface is rebuilt. Motion profiles (Calm, Standard, Vivid), a 4px spacing scale, real letter spacing, a redesigned left rail, Chat as the main screen, and a white theme that follows the system.
• Sources are real rows. Each source shows the site's own icon, the host, the page title and a trust mark. One click opens the page.
• Compare is a banded table with a pinned head, a fixed attribute column, and no sideways scroll.
• The local model store has filters, sort, model tiles, pull progress, and a two-press Remove.
• The browser extension installs with one action and shares the theme and the motion profile.
• The agent operations are rewritten: snapshot-first site recognition, a grounded price, seller questions, and a deep-check source.
• The documentation is rewritten in ASD-STE100 Simplified Technical English, with a new docs/DESIGN.md.

Build

• These builds are unsigned. Windows SmartScreen may warn before the MSI runs. Select More info, then Run anyway.
• The MSI installs for the current user under %LOCALAPPDATA%\Programs\Kriko and needs no administrator prompt.
• The in-app updater installs Windows MSI releases.
• SHA256SUMS.txt holds the SHA-256 digest of both files.

Checksums

See the attached SHA256SUMS.txt.

Kriko 0.1.0

Kriko starts over: this is 0.1.0. The version reset is the release — a re-founded line of numbers for the reworked desktop, merged through PR #218.

• A real icon set. Every navigation icon and control glyph is a professional, consistent set; agent tiles show the actual brand marks (Claude, Codex, Copilot, Gemini, Ollama) instead of generic glyphs.
• Room to breathe. The Local model drawer names each size's quantisation (default q4_K_M), an estimated download weight, and whether it fits this machine's VRAM.
• Sites teach with the flow explained. The Sites screen says what the agent does and what the extension still needs, before it is needed.
• Queue visibility. The Compare nav row carries a badge with the count of products waiting in the research queue.
• A bundled code face. Logs and code render in JetBrains Mono everywhere, no system fallback.

Windows only for this release: install kriko-0.1.0-x86_64.msi, or use the portable ZIP. No macOS packages were built.

Read this if 2.0.0 is installed. 0.1.0 sorts below 2.0.0 on purpose, and the upgrade-aware installer refuses to downgrade: uninstall the newer Kriko first (Apps, or Settings → Danger zone), then run this MSI. The in-app updater will not offer 0.1.0 over a 2.0.0 install. Your ~/.kriko store, preferences and history are untouched by both the uninstall and this install.

The root SHA256SUMS.txt verifies the Windows MSI and portable ZIP. packs.json indexes the two bundled catalogs (cars.kpack, drill.kpack) against this release.

These builds are unsigned: SmartScreen warns on first run (More info → Run anyway).

The complete gate passed on the exact release source: 3,127 Python tests (22 skipped), 269 JavaScript tests, 135 Rust tests, the clean-wheel import/CLI checks, and version agreement. The frozen sidecar answered its handshake; the app stayed up 25 s without panicking; the MSI was installed and every screen of the installed app was photographed (40 captures, including the 940×640 minimum) and the MSI then removed. macOS was not built or observed for this release.

Source: 9089898764cc7d49e7aca692bdf0793d6800a977, the merge of PR #218 into main.