Skip to content

v3.6.0 — All fifteen

Choose a tag to compare

@MarkusNeusinger MarkusNeusinger released this 01 Nov 23:19
· 1589 commits to main since this release
5983830

The window since v3.5.0 holds 253 merged pull requests (#12160 to #12506). The release takes the "Use with my data" agent chat, which only admins can use for now, to all 15 catalogue libraries, with the fixes from the user test of 2026-11-01. The full record is CHANGELOG.md, and each PR carries its own reasoning.

Added

  • plotly and bokeh in "Use with my data", for the admin's live tests. All 15 catalogue libraries are served; each one's 80 % eval matrix gates the public release (#12490).
  • MUI X in "Use with my data", for the admin's live tests. plot.tsx mounts as a React component; a MUI X 7 file is served coupled, never blocked (#12483).
  • Chart.js, D3, ECharts and Highcharts in "Use with my data", for the admin's live tests. They render through the catalogue's browser harness in its sandbox mode (#12479).
  • A Chart.js, D3, ECharts or Highcharts plot downloads as an interactive plot.html. A self-contained page whose plot code runs in an opaque origin (#12480).
  • A Chart.js, D3, ECharts or Highcharts plot plays in the result card. Through a relay, so no token or cookie ever meets the plot code (#12482).
  • Makie in "Use with my data", for the admin's live tests. plot.jl renders in exec mode under the preloaded exec filter (#12464).
  • ggplot2 in "Use with my data", for the admin's live tests. plot.R renders through the renderer's R row in exec mode (#12448).
  • "Use with my data" can adapt and render pygal and lets-plot. 209 of 325 pygal and 284 of 325 lets-plot files are eligible (#12405).
  • "Use with my data" can adapt and render plotnine and altair. 264 of 321 plotnine and 269 of 325 altair files are eligible (#12404).
  • .plot_all() in the admin agent chat: one table, a plot in every library at once. A grid of cards, through the normal run queue (#12417).
  • Agent telemetry: why a run needed its attempts, and with which configuration. The stage that ended every attempt and the score bucket of every round (#12477).
  • The pipeline's Claude steps can run on API credits, by workload identity federation or an API key: the repository variable CLAUDE_AUTH. All 13 Claude steps in the 8 workflows (#12487).
  • Every Claude session's cost in the job summary, and a clean stop when Claude refuses a session. A usage limit no longer uses up a pair's three attempts (#12487).
  • Daily regeneration by implementation instead of by spec: the repository variable REGEN_PAIRS_PER_DAY. The N oldest (spec, library) pairs across the catalogue (#12487).
  • Adopt mode for the implementation pipeline: source_ref on impl-generate.yml and bulk-generate.yml. Code written outside the pipeline reaches main only through the Opus review (#12274).
  • The adaptable-code convention in all 15 library prompts. One replaceable # Data block that ends in a tidy df, one column constant per spec role (#12261).
  • Adaptability checker for Python implementations. core/adaptability.py; the pipeline does not run it yet (#12244).
  • A browser check that the site renders, before and after every merge. scripts/site-smoke.mjs, after a vite bump left every page blank for about 45 minutes (#12236).

Changed

  • The agent chat answers short and asks only when it matters. Questions per reply fell from 0.75 to 0, words per reply from 58.7 to 50.4 (#12506).
  • The agent chat's plot follows the site theme. The first render of each plot version matches the dark or light site (#12505).
  • The agent chat's windows work in parallel; only their sandbox runs queue. At most three runs per user across windows (#12220).
  • The agent's adapter may think and answer up to about 50,000 tokens. 9 of 139 Makie adapter calls had stopped at the old cap of 16,384 tokens (#12475).
  • The agents service runs on 1 vCPU and 2 GiB. Its hourly p99 peaked at 0.46 of a core and 370 MiB (#12456).
  • Python 3.14 everywhere. The project floor, the pinned interpreter (3.14.9), every workflow and the three service images (#12247).
  • MUI X implementations are generated against @mui/x-charts 9.16, React 19 and @mui/material 9.5. Up from @mui/x-charts 7.29.1 (#12243).
  • R implementations now render on R 4.6.1 with ggplot2 4.0.3. All 325 ggplot2 implementations still render in both themes (#12242).
  • The Julia render stack moves to Julia 1.13 and Makie 0.24. Makie 0.24.16, up from 0.21.9 (#12245).
  • Highcharts renders with 13.1.1 instead of 12.6.0. With the fixes for the 2026-09 core security advisories (#12246).
  • Library and runtime versions are read from the pins, never written by hand. Ten of the fifteen hand-written numbers had drifted (#12240).
  • Gemini second opinions run through Vertex AI, and Claude no longer runs agy. Gemini 3.8 Flash for first reviews, Gemini 4 Argon for plans and second opinions (#12226).
  • Gemini tasks as brief files instead of issues. Large work in packages of at most 20 to 25 files, each in a fresh session (#12458).

Removed

  • Google Jules lane. Jules takes no new sessions after 2026-11-11 and shuts down on 2026-12-31 (#12453).
  • The auto-update workflow for PR branches. Its merge commits held each PR's CI at action_required (#12237).

Fixed

  • A follow-up message no longer changes the plot's language. The plot's text keeps a language of its own (#12504).
  • The agent's reply names the texts the plot draws, in all 15 libraries. Before, only matplotlib, seaborn and plotnine reported them (#12503).
  • The agent chat asks before leaving drops the session. A reload, a closed tab or a link off /debug/agent asks first (#12502).
  • A plot can no longer run code of its own as an import. The render guard trusts only the interpreter's own loader (#12484).
  • The site renders again after the Vite 8.1 bump. Vite 8.3.5 emits Rolldown's runtime helper as its own chunk (#12235).
  • Two merges in a row no longer leave production on the older commit. A wait-turn step in all four Cloud Build configs (#12222).
  • Related plots load in about a millisecond instead of about 750 ms. One cached catalogue index, byte-equal on 1,266 combinations (#12214).
  • Currency columns read three-digit comma groups as thousands. "$1,200" reads as 1200, with a warning (#12228).
  • JavaScript implementations get exactly one header. The stale block left the 182 catalogue files that carried it (#12299).
  • Implementation PRs can no longer change dependency files. impl-generate.yml fails such a generation, impl-merge.yml refuses such a PR (#12240).
  • One title and description per page. The shell's defaults step aside while a page renders its own (#12442).

Security

  • The render harness has the kernel refuse every program start. A seccomp filter refuses every execve and execveat (#12410).
  • Every in-process run requires the harness's exec filter. Every render, data run and conversion (#12414).
  • The renderer sandbox stays tight on Python 3.14. Only scipy's import-time ctypes.cast passes the plot guard (#12247).
  • Jules and Dependabot PRs run CI without secrets or an OIDC token. Only Run E2E Tests logs in to GCP (#12429).
  • No CI on pushes to feature/**. Feature branches get CI through their pull requests (#12441).
  • Gemini reads only the owner's issue comments. (#12458)

Full Changelog: v3.5.0...v3.6.0