Anvil logoAnvil
Product Updates & Releases

What’s new in Anvil

New features, engine improvements, model rollouts, and fixes shipped directly to the Anvil platform and CLI ecosystem.

Current Version

v1.32.0

Last Updated

October 27, 2026

Total Releases

65 Updates

System Status

Operational

4 weeks ago
v1.21.0MINOR
Platform Update
Added
  • New anvil-studio skill: a beginner-friendly creative studio that takes an idea to finished posters, social ads, carousels, product shots, app mockups and short promo or motion-graphic videos in six stages (idea, plan, assets, build, review, deliver). It asks at most three questions with starred defaults, offers three genuinely different creative directions before any plan, critiques every render like an art director before you see it, and keeps every stage as editable Markdown in a per-project prompter/studio/ hub with a gallery, revisions, a cost ledger and per-platform exports. Carousels render as one swipeable set with a LinkedIn PDF export; product shots use your own photo, uploaded to kie.ai only after you say yes; --audit <file> critiques a piece made anywhere. Images come from kie.ai with your own key (or a free local fallback), video from the HyperFrames skills.
1 month ago
v1.20.0MINOR
Platform Update
Added
  • davinci now checks the technical details that make a page behave well, not just how it looks. When it builds a high-fi preview, implements a design in your codebase, or audits a live page, it also checks form fields (autocomplete, input types, paste never blocked), long text that must truncate, images with explicit sizes, filters and tabs kept in the URL, touch and safe-area handling, dates and numbers formatted for the user's locale, and hydration-safe markup. Audit reports rank these findings alongside the existing responsive, motion and accessibility ones, and preview messages carry a "Mechanics code-checked" line. The checks are adapted from Vercel's Web Interface Guidelines and ship inside the skill, so nothing is fetched at run time.
Fixed
  • The design-taste-frontend ruleset (and the copy bundled in davinci) no longer tells the agent to write transition: all for medium motion intensity, which animated layout properties the page never meant to animate. It now names transform and opacity explicitly.
1 month ago
v1.19.6PATCH
Platform Update
Changed
  • The project-orchestrator and feature-planner skills now load only the instructions for the mode you are in. status, continue and run no longer load the whole interview, and a new interview no longer loads the run loop. feature-planner also loads its cross-repo rules only when a bound repo owns part of the work, and a new review-planner review no longer loads the run loop.
Fixed
  • feature-planner no longer scaffolds the first increment of a feature while a companion change in a bound repo still blocks it. Previously that left the roadmap stuck: run stopped with an out-of-order error and continue told you to implement the blocked increment. Now the first increment waits, and run or continue handles the companion first.
  • The first run in the feature-planner, review-planner and project-orchestrator skills no longer records tests broken by already-started work as pre-existing failures. When work began before the first run and tests fail, it asks which failures are pre-existing; under --yolo it treats every failure as blocking and says so.
  • feature-planner and review-planner can now be uploaded to claude.ai and the API (their descriptions no longer contain angle-bracket placeholders), feature-planner's English testing guides no longer keep Indonesian labels from the template, and testing guides in feature-planner and project-orchestrator now pick up your project's design system when the skills are installed through the Prompter CLI.
  • review-planner run, continue or status without a review name now resumes your one in-progress review, or asks which when there are several, instead of starting a new review. A mistyped review name now asks which review you meant. Fix roadmaps also consistently put critical findings first when they span several modules.
  • run --review in the project-orchestrator, feature-planner and review-planner skills no longer archives a change the reviewer flagged for human review just because it had no blocking findings. The run now asks whether to archive anyway or stop, asks again if you re-run after stopping, and always stops under --yolo.
  • The project-orchestrator interview no longer asks questions with more choices than the question tool allows; integrations are now asked in three short groups. The Phase 0 proposal it scaffolds marks human-only checks as (manual) so run pauses on them, a re-planned roadmap can no longer reuse a change id left over from the old plan, and English testing guides no longer keep Indonesian labels from the template.
  • davinci can now be uploaded to claude.ai and the API (its description no longer contains an angle-bracket placeholder). When Playwright is missing it installs with your project's package manager (pnpm, yarn or bun when their lockfile is present) instead of always using npm. On designs that serve a proposal or roadmap, the preview message no longer says it will build the design in your codebase — it says the design is recorded for the apply flow. If you keep .preview/ tracked in git, the saved login session for audits behind a login is excluded from commits.
  • davinci now offers to build pages in parallel with subagents when a design has four or more pages and you didn't pass --parallel. Previously the offer almost never appeared. It asks once, either when confirming the page list or with the first page it shows for approval, and doesn't ask again after you decline. In a parallel build it always builds the first page itself for your approval before subagents build the rest.
1 month ago
v1.19.5PATCH
Platform Update
Changed
  • The project-orchestrator, feature-planner and review-planner skills now check whether your AI tool has the anvil-draft, anvil-forge and anvil-audit sub-agents before handing off work in continue or run. If any are missing, they offer to install them with prompter agents: they can run it for you, you can run it yourself, or they carry on with generic sub-agents. Installing keeps the agents you already have for other tools. Unattended run --yolo runs never stop to ask, and carry on with generic sub-agents instead.