v1.1.3 — long skills keep their later steps

Long skills keep their later steps through a long session, a question nobody answered is never taken as an answer, and the shipped permissions approve less. A fix release for 1.1.2.

Highlights

• The 14 longest skills keep their later phases after a compaction. Claude Code re-attaches only the first part of a skill after a compaction, so the later phases of a long skill could drop out mid-session. /adopt, /architecture-decision, /architecture-review, /create-architecture, /design-system, /dev-story, /gate-check, /prototype, /review-all-gdds, /setup-engine, /smoke-check, /story-done, /test-setup and /ux-design now keep each phase in its own file under references/. The steps themselves are unchanged, and /skill-test and /skill-improve read those files too.
• A skipped, dismissed or empty question is not an answer, in any mode: nothing is written, set or picked on it, and it is asked again. Agents return their options as a list for the skill to ask, and CLAUDE.md names the three writes that need no "May I write?".
• Settings read the same from a subfolder. Started in src/, the settings reader missed per-system overrides, the notes about invalid or locked values, and the older review-mode.txt and stage.txt. /settings and /perf-profile work from a subfolder too. A local platform.cert_tier can no longer override the lock.
• Hooks and checks. The commit check works for a project in a subfolder of a larger repository, checks a JSON data file stored with Git LFS through its working copy instead of blocking it, sees files a git add in the same command stages, and stops asking non-system files in design/gdd/ for GDD sections (#133). /smoke-check no longer waits forever on Unreal.
• Every skill opens with a title (#83, thanks @specterslient95-lgtm), and /start writes a single modes: block.
• Smaller fixes: registry lookups no longer match combat-ai when asked for combat; thirteen skills name the minimal route as a next step; team-polish and team-ui keep their core members at lean; /scope-check's verdict bands no longer overlap.

Security

• The shipped allow rules no longer approve deleting a branch or writing files without asking. Four rules matched any text after the command name, so git branch -D, git log/git diff --output=<file> and python -m json.tool <in> <out> ran without a prompt. Those rules are removed; read-only git commands still run without one.
• Start Claude Code at the project root. It loads .claude/settings.json only from the folder it starts in, so a session started in a subfolder runs without the project's hooks, deny rules and ask-first mode. Claude now warns when it was started somewhere else.
• .env files are denied to Edit.

Upgrading from 1.1.2

No setting changes meaning and nothing is migrated automatically. If you merged settings.json or copied its allow rules into settings.local.json, remove the rules listed in UPGRADING.md → v1.1.2 → v1.1.3; a copy keeps approving them. If you edited one of the 14 long skills, take the shipped skill and make your change again in the phase file that now holds that text. If your project.yaml has two modes: blocks, merge them into one.

Tested on

• Platforms: Windows, Linux (bash 5 and 3.2) and macOS (bash 3.2 with BSD tools), with and without jq.
• Claude Code: live skill checks on 2.1.292 and 2.1.293.

Full changelog: v1.1.2...v1.1.3