Skip to content
v1.4.0MIT

Viper's skill stack: viper-mode, how, why, architect, blast-radius, interrogate, follow-through, qa-walkthrough, principle and formal verification skills for Cursor agents.

Changelog

All notable changes to vstack. Versions follow semantic versioning.

Unreleased

1.4.0

Added

  • modern-web-guidance and chrome-extensions, vendored unmodified from Google Chrome's Modern Web Guidance v1.0.0 (4dd9683, Apache-2.0, LICENSE and NOTICE.md in each folder). A short vstack SKILL.md scopes the trigger to web frontend work and looks guides up in local files (INDEX.md, guides/); it never runs the upstream npx ...@latest CLI, its install command, or legacy-scan, so nothing unpinned runs and no telemetry is sent. Google's own SKILL.md is kept as UPSTREAM_SKILL.md (and GUIDE.md for extensions).
  • scripts/sync_modern_web_guidance.py resyncs an exact upstream tag or commit (it refuses latest and branches), blocks on hidden Unicode, skips scripts, and writes VENDORED.json with a SHA-256 for every file; --check fails if any vendored file drifts. Steps: docs/modern-web-guidance.md.

Changed

  • elevate, land-it and qa-walkthrough point web frontend work to modern-web-guidance (no token growth).

1.3.0

Added

  • elevate, a skill that raises UI craft toward premium, modern, fluid feel (hierarchy, type, color, space, motion, states, anti-slop); not a feel diagnosis (/elevate).

Changed

  • Encode reflect 2026-11-13 Accepted lessons: echo unique names (iyow / never-block); prove contrasted install channels; dual-ship + Projects verify (setup-vstack GUIDE); one-card / serialize shared ship files (follow-through / separate); Cursor skills-missing checklist + bag-and-tag three-manifest bump and ship-time skill-dir proof.
  • updater: a saved fetched/ body that does not parse, or has no usable row (a Blizzard table with no us row, an OpenRouter list or crates.io index cut off mid-file, an npm, PyPI, Minecraft, Rust channel or Node index body with no version, an empty tag list), stays under Unchecked sources as "saved body unparseable, truncated?" instead of dropping out of the report or crashing the run, and the report lists each fetched/ file it used with its size. The skill says to save raw bytes only, run the repo's fast checks before opening a bump PR, leave draft with the last code push, and keep the run open until every PR it opened is merged.
  • updater: an advisory listed in a deny.toml [advisories].ignore drops to the watch list and the report names it under "Ignored advisories" with its recorded reason; a source the HTTPS proxy refuses reads "proxy blocked (403)" instead of URLError; bump --area cargo prints the new crates and new duplicate majors a lock refresh pulled in. The skill says to try each blocked source once, try each major as its own small PR before carding it, and name every medium-or-higher finding in the report.
  • follow-through: scripts/closeout.py is the closeout gate. Given every PR a run opened, it refuses "done" unless each one is merged or named --pending with its blocker.
  • AV ship check lessons: av-qa-walkthrough covers apps with no player (reply playback, barge-in echo test, codec boundaries, devices), prefers platform resampling over hand-written DSP, makes each scan line name the command run, keeps the real-media walk as a committed script, and puts decisions only on the decision card; follow-through reads the merge commit and checks when someone else merged your PR.

1.2.2

Fixed

  • Cursor Projects / Cloud Agents packaging: add root Agent Plugins plugin.json alongside .cursor-plugin/ so Cursor-hosted marketplace installs can resolve skills the same way GitHub Import from Repo already does; expand docs/CONSUMING.md Projects checklist (Import from Repo, Required, Auto Refresh, new agent; Sync Skills is personal-only).

1.2.1

Added

  • optimize, a skill for app performance work: measure the real artifact, rank hot spots, fix root causes, remeasure (/optimize).
  • bag-and-tag, a skill for the promote ritual: beta→main, VERSION/CHANGELOG bump, annotated tag, GitHub Release artifacts, installer verify (/bag-and-tag).

Changed

  • setup-vstack installs and configures vstack for the current coding agent (Cursor, Claude Code, Codex, Antigravity, Muse, Windsurf, OpenCode, Grok Build); Cursor-only model rule; see docs/CONSUMING.md.
  • README.md restructured to standard GitHub conventions (badges, table of contents, clearer install and skill catalog sections).

1.2.0

Added

  • land-it, a skill that deep-dives a whole app for product-feel misses (UX that never quite lands, weak interactivity, QoL gaps, code shortfalls) and ranks low-effort fixes (/land-it).
  • Multi-agent install surface: scripts/install_for.sh places skills into documented host paths (.agents/skills, .claude/skills, .grok/skills, Cursor plugin); install.sh gains --target agents and --target grok. Consumer docs (CONSUMING.md, USAGE.md, README) include a Works-with matrix for Cursor, Claude Code, Codex, Antigravity, Muse, Windsurf, OpenCode, Grok Build, and generic AGENTS.md readers; Aider/Continue and GrokHub stay documented-optional / deferred where formats are unconfirmed.
  • updater, a skill that finds every part of an app that needs updating: Cargo, npm and PyPI dependencies, RustSec and npm advisories, Rust and Node toolchains, GitHub Actions, World of Warcraft .toc interface numbers (from Blizzard's version service), Minecraft releases, AI model ids, API hosts, launchers, mod sources and locale gaps. A stdlib script (skills/updater/scripts/updater.py) does the checks and writes a ranked report; bump --area applies the safe ones. Anti-cheat guards are listed but never edited.
  • stageshow, a design skill for faux product UI stills, cursor demos and bento grids, vendored from Stageshow by Benji Taylor (MIT) at upstream 8020574. Its short SKILL.md stays under the 400-token cap and points to the full GUIDE.md.

Changed

  • viper-mode, the always-apply rules, and agent docs now require RPE (Research, then Plan, then Execute) before edits, and waiting only on real (not CI-only) conflicts with other open work.
  • docs/USAGE.md consumer install matches docs/CONSUMING.md (plugin / install_for.sh / install.sh; no stale cp -R or missing-snippet guidance).
  • docs/ci-policy.md lists which CI jobs run on which event. The nightly CI run skips a commit that already passed CI, and every job has a timeout.

Fixed

  • token-check no longer fails every push and nightly run: pushes measure growth against the previous branch tip and read the approval label from the merged PR, and runs with no base check only that tokens.json is fresh. The job drops its pip cache, which could stall for 10 minutes, and times out after 5 minutes.

1.1.0

Added

  • 16 workflow skills ported from pstack: how, why, teach, recall, architect, figure-it-out, show-me-your-work, blast-radius, interrogate, tdd, correct, reflect, swarm, benchmark-checklist, technical-writing, typescript-best-practices.
  • 18 principle-* skills, one for each principle viper-mode cites.
  • follow-through (task ledger, stall detection, proven done) and av-qa-walkthrough (audio and video QA).
  • Compressed statechart and ketchup-video ports.
  • Release packaging: install.sh, scripts/package_release.sh, and a release workflow that publishes tarball and zip downloads with SHA-256 checksums.

Changed

  • viper-mode restates the goal and smallest change on the first message, then routes to the new skills.
  • ketchup-video adds a subtitle strip and measured slide holds.
  • Synced to upstream pstack c742dd6.
  • Agents can load viper-mode, the principle skills, and the core workflow skills on their own; you no longer need a slash command for each.

Fixed

  • how and why frontmatter now parses as YAML.

1.0.0

  • First release: viper-mode, setup-vstack, in-your-own-words, sharpen, unslop, ketchup, qa-walkthrough, verification and formal-verification skills, principle-fix-root-causes, and Cursor and Claude Code plugin packaging.