v1.6.1

Fixed

• trp — Phase 0 grounding now runs three mandatory checks before writing the package,
  closing two failure modes surfaced by a real ticket where a bug reported against QA got a
  first-pass package proposing to cherry-pick a fix that was already deployed to that same QA
  environment. First, a Timeline check enumerates every candidate fix or prior related commit
  discovered while grounding as {SHA, subject, branches, merge date per env} and rejects any
  candidate whose merge date on the reproducing env is on or before the ticket's report date —
  it is already deployed, so it cannot be the answer. Second, an Evidence-sufficiency check
  forces the mode: if the ticket describes runtime behaviour the repo alone cannot explain (a
  thrown error class, an HTTP response body, browser/localStorage state at the failure moment,
  environment-only reproduction), the deliverable is a runtime repro plan and the package is
  spike-solve, not a solve package with hypotheses wrapped in caveats. Third, ranked
  hypothesis lists with a "recommended default" are banned in a solve package — that content
  now has a physically distinct spike-solve template variant with a Hypotheses-to-test
  section, a Phase 1 spike plan, and DEFERRED Phase 2+, so a spike deliverable cannot pass as
  a fix proposal. gates.md's failure catalogue extends the "Suspected/likely" ban to those
  ranked lists and adds a bullet for the already-deployed-fix failure mode.

v1.6.0

Changed

• trp — the heavy phases now run in Agent-tool subagents pinned to the skill's model
  while every gate stays in the main conversation: a read-only Phase 0 grounding subagent
  returns the Full TRP Package, the approval gate fires inline, a delivery subagent runs
  Phases 2 through 4 with the gates carried verbatim in its prompt, and the pr-review loop
  plus Phase 5 stay inline. A context: fork skill would lose AskUserQuestion, so the
  gate could not live inside a forked run; the reference files and the operator playbook
  now say the only subagent before approval is that read-only grounding pass.
• sec-audit — the audit sequence (resolve, threat-model, scanners, live probes, scoring,
  advisory drafting, lint, coverage claim) runs in one subagent that stops at the first
  hard human-gate and returns the decision needed as options with a default; the picker,
  preflight, target and mode confirmation, the multi-agent opt-in, and all six gates stay
  inline, and the same subagent is resumed with the operator's decision.
• copy-audit — the script phases run in subagents (extract + bundle-emit; apply-verdicts
  with a compact before-and-after summary; apply + verify on the approved files only) while
  the Workflow review fan-outs and the review-before-apply gate stay inline.

v1.5.2

Fixed

• pr-review / trp — the CLI entrypoint guard compared import.meta.url (resolved through
  a symlink) against process.argv[1] (unresolved), so under the standard ~/.claude/skills
  install layout main() never ran and the scripts exited 0 with no output and no file
  written. fetch-pr.mjs, fetch-ticket.mjs, and clickup-update.mjs now resolve argv[1]
  with realpathSync before comparing.
• pr-review — Azure DevOps marks some rename and add-shaped-as-edit change entries with a
  changeType that omits "add", so the loop fetched the old blob at the base commit and died
  on the 404. The blob helper now separates real 404s from other transport errors, and the
  loop follows sourceServerItem / originalPath for the pre-rename path and treats a
  missing base blob as an add (or a missing head blob as a delete) — with the effective kind
  recorded in status — instead of refusing the whole PR.

Changed

• pr-review — a URL invocation now runs in a subagent, so a 30 KB diff and the finding /
  self-verify loop never share the operator's context. The subagent's final message is the
  paste-ready block, verbatim. render-review.mjs wraps its own output in an outer fence
  longer than any inner run so a suggestion body's triple backticks copy cleanly through
  the subagent boundary.

v1.5.1

Fixed

• copy-audit (comment mode) no longer deletes a comment that is the sole content of an
  otherwise-empty block. Removing it empties the {}, which reads as unintentional and
  trips eslint no-empty / no-empty-function (observe() { /* no-op */ },
  () => { /* stub */ }). apply and the shared splice now refuse such a delete and keep
  the comment.

v1.4.0

Added

• Every skill now has an animated demo on the marketing site. The "Watch each skill run" section
  plays a short, synthetic Claude Code session for pr-review, sec-audit, trp, and copy-audit — the
  typed command, the streamed output, and the final verdict — two per row, then settles on the full
  transcript. The demos are self-contained and theme-aware (no rasterized screenshots), play only
  under prefers-reduced-motion: no-preference, and add no JavaScript; a build-time schema
  validates every transcript.

Changed

• Rewrote the four skill descriptions, on the site and in the plugin, to be benefit-led and plain.

v1.3.2

Fixed

• copy-audit's holistic reviewer now flags two defect classes it had been missing: dead
  cross-references (a "see X" pointer, a [..](#anchor) link, or an "in this skill's reference
  files" phrase whose target is absent from the corpus) and duplicate or divergent descriptions
  of the same artifact (a file, script, command, or config key described two ways). It had
  passed a reference to a section that was never written and a workflow file listed twice with
  different descriptions.

v1.3.1

Fixed

• copy-audit's per-unit reviewer was too lenient — a blanket keep-bias suppressed flagging as
  well as rewriting, so it passed almost every unit. It now flags every defensible content-pillar
  violation while keeping auto-applied rewrites conservative, which surfaced roughly three times
  as many issues on the same corpus (including a corrupted sentence in a reference doc).
• The copy-review workflow crashed at the workflow runtime's 4,096-item return cap on a
  whole-repo sweep; it now returns verdicts in chunks, with the per-agent journal as the fallback.
• markdownlint only checked root *.md, so every skill and reference doc went unlinted. The gate
  now lints all authored markdown (vendored, build, and toolchain trees excluded), and 173 latent
  violations across 22 docs were fixed.

Changed

• Repo-wide copy pass: tightened 114 spans across the operator playbooks, skill reference docs,
  the READMEs, and the website — splitting run-ons, cutting empty intensifiers, fixing comma
  splices, and repairing a mis-described token fallback and other defects.

v1.3.0

Added

• copy-audit skill — audit product copy, UI microcopy, and (with --mode=comments)
  code-comment slop and test-runner names across a git repo or diff. AST-based extraction
  (tree-sitter for 40+ languages plus babel/Astro/remark/yaml) reaches markdown prose,
  headings, and alt-text, JSON/YAML copy values, and UI microcopy in HTML/Astro/Svelte/Vue/
  JSX/TSX; parallel subagents judge each unit with full-file context against four cited
  content pillars (plain language, inclusive, UX microcopy, voice/grammar) or, in comment
  mode, comment-quality plus Rule 9. Verdicts are keep/rewrite/flag/delete and every rewrite
  is a SHA-guarded char-offset splice that changes text but never code structure, with a
  verify phase asserting zero code-line changes. Read-only until you approve the rewrites.
  Subsumes the standalone comment-audit skill. Runs against a git repo/diff, against pasted
  text or a standalone file with no git (--input / --stdin --as, verified by exact
  pre-image reconstruction), and against PDFs (text pulled with Node's built-in zlib and
  reviewed read-only — never spliced back into the binary).
• Marketing site: a fourth skill card, a copy-audit mockup, FAQ entries, and an Operator's
  Playbook PDF for copy-audit.

Changed

• The plugin now bundles four skills; the plugin and marketplace manifests, the READMEs, and
  the marketing site update from "three" to "four".

v1.2.3

Features

• Generate release notes from commits with git-cliff (release)

Documentation

• Rewrite the README to lead with the product

v1.2.2

• Added a curated CHANGELOG.md and a VERSION file, kept in lockstep with plugin.json and the release tag: a gate rejects any tag or version bump that does not match, and release notes now come from the matching CHANGELOG section instead of the tag message.
• The site's Resources "Changelog" now links to CHANGELOG.md.
• Hardened the accessibility test to settle animations and fonts before scanning, removing an intermittent contrast flake.