Add a repo-scoped Claude Code skill at .claude/skills/analyze-action-pr/
that triages allowlist PRs end-to-end: extract action refs via
verify-action-build --from-pr, classify each finding (pipe-to-shell,
unverified-download, nested-action-issue, verify-script-bug, all-clean),
look up what verification material upstream already publishes, and
draft the right next moves — recommend approval, open an upstream issue
+ ping the PR author, or fix verify-action-build itself with a
regression test.

Includes an "Improve this skill" step so future runs that hit a new
shape (false positive, novel verification mechanism, asset-naming
convention) leave the skill better than they found it: the SKILL.md
grows by a row in the case table or a line in the precedents table,
not by re-deriving the analysis.

The skill encodes shapes recently triaged: PR apache#795 / apache#798 (TS-generic
regex hole), PR apache#802 (carabiner-dev nested install actions), PR apache#803 /
apache#804 (multi-action extraction), PR apache#806 (jbangdev pipe-to-shell).

Replace the broad .claude/ rule in .gitignore with specific entries
for personal state (settings*.json, worktrees/, debug/) so .claude/
itself can host shared assets without leaking per-user files.

Pre-commit on PR apache#807 failed because .claude/skills/analyze-action-pr/
SKILL.md was missing the Apache license header.  The insert-license
hook auto-fixes locally but aborts CI because there's no follow-up
commit to land the change.  Add the header so CI passes.

Document the local pre-commit workflow in AGENTS.md so future agent
runs install prek and run ``prek run --all-files`` before pushing,
catching the same class of bounce locally instead of in CI.

Case E previously read as "the script has a bug" — too narrow.  Real
gaps over the life of verify-action-build include whole-action-type
support (Deno, Dart), new build flows (npm run start, multi-step
Docker), new verification heuristics (sibling sha256sum, JSON data
fetches, *Json helpers), false positives on legitimate inputs
(multi-stage FROM, vendored node_modules), and extraction-shape gaps
(hashes under an existing actions.yml key).

Reword the case E row to cover all of these, and add a sub-table
under step 5/E that lists 15 historical extensions with the PR or
commit that closed each.  The table is the template a future agent
reaches for when "the verdict is wrong but not for a security
reason" — instead of waving the finding off, they can match it to
the closest precedent and propose a similar extension.

Also tighten the steps: name each module under
utils/verify_action_build/ and what it owns (pr_extraction, security,
verification, docker_build, action_ref/release_lookup/github_client),
so picking the right place for the fix isn't a treasure hunt.  Add
an explicit "run prek before push" step pointing at AGENTS.md.