Skip to content

v4.1.0 — Fine-tuning

Choose a tag to compare

@MarkusNeusinger MarkusNeusinger released this 12 Nov 06:09
· 1029 commits to main since this release
e13735c

The fine-tuning after the public launch. The agent chat ("use with my data") gets one message box for data and text, a data step that derives columns, and a thread that says what each turn changed. The stats page counts machine readers and shared links, the review pipeline starts recording what it used to assume, and all 323 MUI X implementations now run on MUI X 9. 98 PRs (#12546 to #13058), plus 17 Dependabot bumps and 406 pipeline PRs; the full record is CHANGELOG.md, and each PR carries its own reasoning.

Added

  • Derived columns in the data step. compute, diff, rolling and blank; in the persona tests, five of ten personas had got no plot because a column could not be derived (#12597).
  • Text operations in the data step. split, extract, strip and replace, with literal texts only, never a pattern (#12556).
  • Dates as numbers in the data step. The calendar year, the days since the earliest date, Unix seconds or a decimal year beside a date column (#12558).
  • The top N rows, the last value per month, over several columns, and several values into one category (#12597).
  • The chat rebuilds from your original table. "The first 20 after all" after a top 10 is one call from the original, stored as the next version (#12600).
  • The chat names what it cannot do. A change no operation can make gets a fixed sentence and is counted as a capability gap (#12597).
  • .example_data() in the agent chat. One click asks the assistant for a made-up table to try the plot (#12553).
  • Questions in the agent chat are a turn kind of their own. A message that changed nothing is marked ?, by the server's word (#12826).
  • A plot card names the remembered decisions it kept. kept: y = 7 line · x label · new: log scale on y (#12826).
  • One working line at the foot of the agent chat, with the plot's clock. drawing your plot · 0:42, the try, and the place in the queue (#12950).
  • A binding warning with counts when a plot's key repeats. For the 90 specs whose plot draws one value per key (#12824).
  • The chat header shows the plot it adapts, and .use_my_data() → on the spec hub and in the code tab (#12549).
  • Machine readers on the stats page. Per day, the verified requests of crawlers and AI assistants, split by purpose, with the MCP requests beside them (#13015).
  • Machine-read lines: one content-free record per crawler, AI assistant or MCP read. Vendor, purpose, Cloudflare's verdict and route, never an IP (#12962).
  • "Shared links" on the stats page. The links shared over the last 28 days by day, platform and plot page, from counters that keep no address and no user agent (#13055).
  • The review warns when the staged renders don't come from the code under review. A sha256 of the rendered implementation is staged beside the PNGs and compared at review (#13034).
  • The review records whether an implementation keeps the generator's static contract. Six rules, record-only (#13045).
  • The pipeline records when a new version stops working in the agent chat. Record-only, in generate and repair (#13054).
  • A project skill audits the live catalogue and repairs what the owner picks, and a project skill triages the agent chat's feedback cases in rounds (#12825, #12551).
  • A docs page on how anyplot is built with AI agents. Who does what, the path of a change from idea to production, and the limits of the approach (#12559).

Changed

  • One chat field for everything. The message box takes a pasted table or an opened or dropped file as an attachment, and Enter sends it (#12565).
  • One click from checked data to the plot. .confirm() is gone; .create_plot() confirms the data and starts the plot (#12557).
  • The plot button sends, reads and plots in one click (#12833).
  • The agent chat is one terminal-style window, and the agent chat reads in chapters, and the data card stays open (#12587, #12604).
  • One chat page for visitors and admins. /agent for everyone; the admin's chat looks and behaves like a visitor's (#12589).
  • The agent chat stays English until you pick another language. Another language is offered once, in one line (#12865).
  • The assistant's replies are shorter and never walk through the page. One or two short sentences with the counts that matter (#12576).
  • The plot code draws every row it gets, in all four languages. Grouping, pivoting and filtering the user's rows left the plot code (#12953).
  • A plot of rows that repeat a key combines them first, as a data step. A numbered, undoable data version, made before the plot runs (#12953).
  • The chat assistant's turn bounds are code, not prompt sentences. A cap per tool and turn; the call after it is refused with a fixed message (#12853).
  • "plot the new data" keeps what you tuned when it can (#12834).
  • One column can play several roles in the data step. The y and the size of a bubble, a value and its label (#12564).
  • Plot votes count per version of the image. A regenerated plot counts from that moment on; nothing is deleted (#12830).
  • The gallery search ignores generic words and accents (#12608).
  • The layout sends every page's pageview, at the navigation (#13046).
  • robots.txt welcomes every crawler, Bytespider included, and AI_AGENTS names 18 more crawler tokens and a new seo purpose (#12868, #12962).
  • Every release runs the agent evals first. Prep and scope always, the matrix smoke when the window changed the plot pipeline (#12579).
  • The agents' tree-sitter grammars are baked into the image, and the service never downloads one (#12943).

Removed

  • Core Web Vitals no longer go to Plausible. They were 44 % of the billable events, 17,044 of 38,818 in 28 days (#12860).
  • The Plausible bot site bots.anyplot.ai. It saw about 1.7 % of real crawler reads; the machine-read lines replace it (#12962).
  • og_image_view no longer goes to Plausible. It carried the fetcher's address and user agent there and measured little (#13055).
  • The public GET /agent/session/bundle. No page has called it since the public chat dropped .save() (#12555).

Fixed

  • A merge whose renders never reach production now fails instead of passing. #12800 had been promoted with nothing copied (#12819).
  • A repair can no longer send unchanged code back to review, and a generation whose code was never pushed fails (#13034).
  • A pair the regen gate kept on a regression or a technical failure comes back after the attempt window. It used to queue behind about 4,700 pairs (#12605).
  • A cut chat stream no longer hides a plot the server made (#12572).
  • A long reply is shortened, never refused (#12590).
  • A web table copied one cell per line, and an inline list such as a:7,b:12,c:5, are data now (#13019).
  • A web table with grouped columns arrives whole, and a pasted Markdown table keeps its column names (#12574, #12602).
  • An id no longer labels the bars, and a founding year is no longer the measure (#12571).
  • Prices no longer bind as a MACD line or an RSI, and raw measurements no longer as control-chart summaries (#12612).
  • Pictograms, waffles, dot matrices and parliaments no longer read as made-up data (#12552).
  • The site's preview images keep a plot's background instead of turning it black (#12863).
  • The chat recognises a library however you write its name (#12855).

Security

  • Cloudflare Turnstile on the calls that open or store something for a visitor. Opening a public chat session, a chat feedback case and a site feedback entry needs a token, and the check fails closed (#12555).

Full Changelog: v4.0.0...v4.1.0