What's New (Upgrade & Change Research)

On-demand research skill. For a named target -- or, with no target, the things
the current repo depends on -- find what changed between what is in use and
the latest, and summarize it as: breaking changes, deprecations, new
features/capabilities, and notable fixes.

Fetch changes programmatically, not by reading rendered web pages. Versions,
tag lists, changelogs, and service announcement streams have machine endpoints
(JSON, RSS/Atom, git) that return small structured output.
skill://whats-new/references/recipes.md is the cookbook. Reserve web fetching
for genuine prose (a migration guide, a post explaining a breaking change).

Target kinds

• Versioned software (library, CLI, framework, runtime, package, container
  image): has a registry + a git repo + semver. Research a version span
  current→latest. Recipes steps A--D.
• Service / technology / platform / API / model family (AWS Bedrock,
  Anthropic/Claude models, OpenAI, GCP, Azure, Stripe, GitHub Actions): no
  semver. Research a dated announcement stream since the user's reference
  point. Recipes step E.

A target can be both -- research both sides when relevant.

Inputs

Honor whatever the user supplies; only discover the rest.

• Target -- if absent, discover it (step 1).
• Reference point -- for software, the current version; for a service, the
  user's baseline date or what they currently use. If absent for a service,
  default to the last 6 months and state it.
• Latest -- the registry's latest version, or "now" for a service stream.

Workflow

1. Resolve the target and its kind. If named, use it. Otherwise run the
   version_gap_scan tool to list declared dependencies. If it returns
   more than one equally plausible candidate, list them and ask.
   Decide: versioned software or service/stream.

2. Resolve sources programmatically.

   • Software (recipes A/B): current version (prefer lockfile), registry
     latest, version list, upstream repo URL.
   • Service/stream (recipes E): vendor's machine-readable change source
     (RSS/Atom feed, release-notes JSON, models API, or changelog repo) and the
     time window.

3. Gather changes programmatically.

   • Software (recipes C): bare git clone -- CHANGELOG at the target tag +
     conventional-commit-classified log across the span. Enrich with host
     release notes via its API.
   • Service (recipes E): pull the feed/release-notes/model-list, filter to
     the window, group.

4. Fill prose gaps only if needed (recipes D): fetch a migration guide or
   deprecation timeline. Prefer vendor guides over third-party blogs.

5. Summarize into the template. LOAD skill://whats-new/references/report-template.md and
   fill every section. Cite a source (release tag, commit SHA, CHANGELOG
   heading, feed entry date, doc URL) for every change flagged breaking or
   touching a public API.

6. Save or return -- return inline; save to a file only if asked or if the
   report exceeds 500 words.

Steering

• Report, don't upgrade -- enforced by the report-only-gate extension.
• Programmatic over manual. Catch yourself reading a rendered registry page → stop and use the matching recipe.
• Cover the whole span. Every intermediate version (software) or the full window (service) -- not just the endpoints.
• Classification is heuristic. A feat:/fix:/! prefix is a signal, not ground truth; read the diff for load-bearing changes.
• State coverage honestly. Name the sources that ran and those that were missing.
• Services have no semver. Anchor on dates, not version numbers. Note announcements ≠ regional/account GA for service targets, always.

Scripts

Script	Purpose
version_gap_scan tool	No-network enumeration of declared dependencies + pinned versions across ecosystems.

For everything network-facing use the commands in skill://whats-new/references/recipes.md.