Lingua v1.3.0

• Scratchpad now keeps live values and failures beside JavaScript, TypeScript, and Python source. Top-level expressions render inline by default, thrown values use the error treatment without losing their full console detail, and one failed capture no longer prevents later expressions from reporting results. Python selects expressions with CPython's AST inside Pyodide so multiline syntax, indentation, declarations, control flow, function bodies, and docstrings retain their normal semantics; Settings and per-tab controls remain available when a quieter editor is preferred.

Read v1.2.0 notes →

• The Cron Parser now writes expressions from plain words. Describe a schedule the way you would say it — *cada 3 días a las 8am*, *weekdays at 9:30*, *every 15 minutes between 9 and 5* — and the utility generates the matching five-field expression, feeds it straight into the existing explanation and next-runs preview, and shows every default it filled in as a visible chip. It understands English and Spanish interchangeably, warns about the classic traps (day-of-month steps restart at month boundaries; date-plus-weekday is an either/or in cron), and refuses honestly when a phrase asks for something a single cron expression cannot express — a continuous every-two-weeks, or two times with different minutes — instead of shipping an approximation that fires when you did not ask. It also reads the vocabulary the established open-source parsers use — yearly and quarterly cadences, bare ordinals like *every 25th* or *the 1st and 15th of the month*, *every other day*, and *first Monday of the month* / *last Friday of the month* via the Quartz nth-weekday operators, each flagged as non-POSIX so you know to check your scheduler. It reads *cada quincena* / *quincenal* / *twice a month* as the 1st and the 15th — the payroll cadence that has no literal cron form but does have a conventional one — and says both that it made that reading and that the resulting gap is uneven. The same grammar is available outside the app as lingua utility cron-phrase, which reads a phrase on stdin and writes the bare expression, with --option annotate=true to append the assumptions and caveats as crontab comments. For free-form phrasing beyond the built-in grammar, an optional assist sends the phrase to the local AI endpoint configured in Settings → AI, validates the model's answer through the same parser before it can land, and suggests lightweight local models (around 0.4–1.3 GB) that handle the task well.

Read v1.1.0 notes →

• Workspace icons now sit on one deliberate density scale. The chrome had drifted into six different icon-button sizes and five glyph sizes, with glyphs floating at roughly a third of their buttons — which is what read as unpolished. Icon-only controls now share two tokened steps (28px buttons with 16px glyphs, 24px with 14px for dense spots like tab close), the editor sidebar toggle no longer fights its own size override, and a build-time guard keeps future surfaces from inventing a seventh size.
• The console's payload-kind chips became a single Types menu. The old row of six chips took nearly half the filter bar and looked identical to the severity chips beside it while behaving inversely (severity chips opt in, kind chips opt out). The menu lists every kind as a checked item, carries a count badge when something is hidden, and opens clear of the panel edges. The console header also drops a duplicated caption that wrapped to five lines at narrow widths.
• Command palette results now rank actions first. Typing a verb surfaces the matching action ahead of snippets and templates that merely contain the word, with label matches boosted within each group.
• The HTTP response empty state told users to click a button that does not exist. It said Send; the button is Run. Corrected in both languages — the Spanish copy also said Enviar while the button says Ejecutar.

Read v1.0.2 notes →

• The HTTP networking library inside the desktop app moves to its patched release: the desktop main process bundles undici to drive the SSRF-guarded HTTP proxy, so its version ships to users even though the package is declared as build tooling and the production dependency audit cannot see it. It now resolves to the advisory-fixed 7.29 line. The hono server library reached by the desktop local MCP server moves to its patched 4.12 release on the same pass, and the build-only fast-uri and postcss toolchains follow. The independently locked license and update Workers move to the same patched hono, undici, and postcss releases, so no deployed surface is left on an advisory line. No behavior changes; this is a dependency-only patch.

Read v1.0.1 notes →

• One workspace now spans experiment, inspection, and project work: write and run six languages, pause JavaScript/TypeScript or native Python/Go/Rust code, model multi-step ideas in reactive notebooks, and move into HTTP, SQL, or focused developer utilities without rebuilding context in another app.
• Desktop project tools are explicit about trust: the integrated terminal, detected test runner, dependency actions, filesystem bridge, and local MCP server all start from an approved project capability, use bounded or filtered process boundaries, and explain where host permissions begin instead of pretending local code is sandboxed.
• Web and desktop share the same local-first product model: browser-capable workflows stay inside workers and local storage, while host-only features are labeled desktop-only and cross narrow typed IPC bridges. Privacy-sensitive actions remain opt-in, telemetry stays off by default, and AI sends only a payload the user previews.
• Repeatable work has first-class artifacts: Run Capsules and bounded multi-file Capsule Workspaces, named HTTP pipelines, notebook execution ledgers, project test evidence, CLI JSON output, and shell completion turn successful experiments into inspectable flows that can be replayed or automated.

Read v1.0.0 notes →