Agent Skills’ frontend skill stops installing test runners just to check its work

In Addy Osmani's Agent Skills 0.6.16, the frontend UI engineering skill stops installing a test runner or type checker to check its own work. It runs the project's own checks instead, so fewer agent runs stall on installs and time out.

Across eleven models in Claude Code, runs that installed a package fell from about a quarter to four percent. Timeouts fell from 53 to 36, out of 198 runs. The old skill did worse than no skill. The new one removes that harm.

The spec-driven development skill now keeps ideas for later in a deferred, out of scope list. When a spec already exists, the spec command writes a small delta spec for the change, instead of overwriting the original.

SPEC.md                    V1, shipped
## Out of Scope (Deferred)
- PostgreSQL instead of SQLite
SPEC-postgres-migration.md  the change, on its own

/build auto now builds the delta spec that is not yet implemented. Each plan records the spec it came from. If a plan came from a different spec, it stops and asks.

The documentation and ADRs skill now judges a stale doc from git history: the last commit to the code against the last commit to the doc. When that history is unordered, it says it can't tell, never current.

last commit to the code  vs  last commit to the doc

stale: code changed after the doc
current: doc changed after the code
can’t tell: rebase, divergent branches, shallow clone

The constraints skill's dependency cruiser gate printed its usage and exited zero, so it always passed. Now it cruises the source folder and fails on a circular import.

$ depcruise --validate src
Usage: depcruise [options] [files-or-directories...]
exit=0
$ depcruise src
  error no-circular: src/a.js → src/b.js → src/a.js
exit=1

The security skill's session example started fine without a session secret, then failed every request with a server error. Now it throws at startup and names the missing variable.

Agent Skills 0.6.16 was released November 14, as a plugin for Claude Code and Codex, and through the skills command line tool for other agents. Only the frontend fix was measured on models.