도구·권한·스킬 표현 정교화 1차 반영

- 권한 요청 카탈로그를 bash/powershell/web_fetch/mcp/skill/file_edit/file_write/git/document/filesystem 수준으로 세분화했습니다.

- 도구 결과 카탈로그에 approval_required, partial, follow-up hint, attention 메타를 추가해 후속 renderer 고도화 기반을 마련했습니다.

- 스킬 갤러리에 모델, 추론 강도, 실행 컨텍스트, 에이전트, 모델 호출 비활성화, 추천 상황을 표시하도록 확장했습니다.

- README, DEVELOPMENT, parity plan, regression prompts 문서를 2026-04-06 11:52 (KST) 기준으로 갱신했습니다.

- 검증: dotnet build src/AxCopilot/AxCopilot.csproj -c Release -v minimal -p:OutputPath=bin\\verify\\ -p:IntermediateOutputPath=obj\\verify\\ (경고 0 / 오류 0)

docs/AX_AGENT_REGRESSION_PROMPTS.md

@@ -108,3 +108,21 @@
   - Cowork 변경: 3, 4, 7, 8
   - Code 변경: 5, 6, 7, 9, 10
 - 체크 후 문서 이력에는 “어떤 묶음을 확인했는지”를 간단히 남깁니다.
+## Tool / Permission Follow-up
+
+11. 권한 거부 후 재시도
+- 프롬프트 순서:
+  - `src 폴더에서 설정 파일을 수정해줘`
+  - 첫 권한 요청은 거부
+  - 같은 작업을 다시 요청
+- 확인:
+  - `reject`와 `approval_required`가 같은 결과 카드처럼 보이지 않음
+  - 재시도 시 권한 메시지와 도구 결과가 중복되지 않음
+  - `bad-approval-flow`
+
+12. 부분 성공 / 후속 안내
+- 프롬프트: `여러 문서 파일을 한 번에 읽고 요약해줘`
+- 확인:
+  - 일부 실패가 있으면 `partial` 계열 안내가 보이는지
+  - 후속 안내 문구가 단순 실패와 다르게 보이는지
+  - `status-noise`

docs/DEVELOPMENT.md

@@ -4941,3 +4941,6 @@ ow + toggle ?쒓컖 ?몄뼱濡??ㅼ떆 ?뺣젹?덈떎.
 - Document update: 2026-04-06 11:27 (KST) - At this point preview, file browser, worktree chooser, and permission-mode popup surfaces all share nearly the same visual language; remaining work is minor spacing/color polish rather than another behavior or structure change.
 - Document update: 2026-04-06 11:34 (KST) - Added `FooterChipBtn` to `ChatWindow.xaml` and aligned the footer work-bar buttons (`권한`, `Git 브랜치`) to the same rounded border, padding, and hover language instead of mixing outline and ghost button styles.
 - Document update: 2026-04-06 11:34 (KST) - This is a footer polish follow-up after the structure split, aimed at making the bottom work bar feel like one coherent tool row rather than a set of unrelated controls.
+- Document update: 2026-04-06 11:52 (KST) - Expanded `PermissionRequestPresentationCatalog.cs` into an action-level permission taxonomy (`bash`, `powershell`, `command`, `web_fetch`, `mcp`, `skill`, `question`, `file_edit`, `file_write`, `git`, `document`, `filesystem`) and added `ActionHint`, `Severity`, `RequiresPreview` metadata for later renderer refinement.
+- Document update: 2026-04-06 11:52 (KST) - Expanded `ToolResultPresentationCatalog.cs` with richer result taxonomy and post-result guidance. AX now distinguishes `approval_required` and `partial` in addition to `success / error / reject / cancel`, and each result carries `FollowUpHint` and `NeedsAttention`.
+- Document update: 2026-04-06 11:52 (KST) - Extended `SkillGalleryWindow.xaml.cs` to expose runtime-policy metadata (`model`, `effort`, `execution context`, `agent`, `disable model invocation`, `when to use`) so AX skills are easier to inspect and maintain like `claw-code` bundled skills.

docs/claw-code-parity-plan.md

@@ -19,6 +19,8 @@
 - Formalized the regression ritual in `docs/AX_AGENT_REGRESSION_PROMPTS.md` by adding failure classes (`blank-reply`, `duplicate-banner`, `bad-approval-flow`, `queue-drift`, `restore-drift`, `status-noise`) and required prompt bundles per change area.
 - Updated: 2026-04-06 10:07 (KST)
 - Continued the maintainability track by moving topic preset rendering, custom preset context menus, and topic-selection application flow into `ChatWindow.TopicPresetPresentation.cs`. This reduces mixed preset UI logic inside `ChatWindow.xaml.cs` and keeps the main window closer to orchestration-only responsibility.
+- Updated: 2026-04-06 11:52 (KST)
+- Continued the tool/permission/skill sophistication track by expanding AX presentation catalogs toward `claw-code` specificity. `PermissionRequestPresentationCatalog.cs` now models action-level permission kinds plus severity/action hints, `ToolResultPresentationCatalog.cs` now models `approval_required`/`partial` result states plus follow-up guidance, and the AX skill gallery now exposes runtime-policy metadata that was previously hidden.
 
 ## Preserved History (Summary)
 - Core loop guards and post-tool verification gates are already partially implemented.