Release 0.3.0

CHANGELOG.md

@@ -2,6 +2,60 @@
 
 ## [Unreleased]
 
+## [0.3.0] — What it could not see, and what it could lose
+
+Two things ran through this release. The app was losing work it had promised to keep, and it was answering questions about files it had never actually looked at. Both failed the same way: quietly, with a successful-looking answer.
+
+On top of that, three new things: it talks to more models, extensions can do considerably more, and a proposal now appears in the code rather than only on a card.
+
+### Your unsaved work stops disappearing
+
+Every one of these was reproduced in the running app before it was fixed, and each is a place where edits vanished with no warning and no undo.
+
+- **Files deleted or renamed outside the app** took your unsaved buffer with them. They are kept now, and you are told they are orphaned so you can decide.
+- **Quitting** with unsaved files did nothing visible: `beforeunload` cancels `app.quit()` in Electron with no dialog at all. There is now a real prompt with save / discard / cancel — and "save" that does not finish does not quit.
+- **Switching projects** dropped unsaved edits without asking.
+- **Discarding in git** only half-discarded, and the delete confirmation promised something it did not do.
+- **Undo of an agent run** reloaded every open model from disk, including files the run never touched — while the screen said unsaved edits would be left alone.
+
+### Files it could not see
+
+- **Anything nested deeper than 8 levels was invisible** to the tree, to search, and to replace — all three walked with the same cap. `packages/app/src/features/x/components/y/z.ts` is already 8. The limit is 16 now.
+- **Dot-directories were hidden wholesale.** `.vscode/settings.json`, `.claude/agents/*.md` and friends appeared in no listing and matched no search; `.github` had been carved out as a special case, which was the tell. What is hidden is now decided in one place.
+- **Neither the tree nor search said when it had stopped early**, so "no results" and "never looked" arrived as the same answer. Both now say which it was — including to the agent, which had been told nothing at all.
+
+### Files it could destroy
+
+- **Opening a non-UTF-8 file corrupted it.** Everything was decoded as UTF-8, so a UTF-16 file grew from 16 to 20 bytes on open-and-save alone, and a CP949 Korean file turned to replacement characters the moment you edited one line. Such files are now refused, with an explanation.
+- **Replace-across-files did the same thing**, on files you never opened — and reported success. Non-UTF-8 files are skipped and named.
+- **A UTF-8 BOM was stripped by opening the file.** Monaco keeps the BOM outside the model, so the baseline never matched and "save all" rewrote untouched files.
+
+### More models
+
+Gemini, and any OpenAI-compatible server you point it at — Ollama, LM Studio, llama.cpp, or a corporate proxy. The address is a setting rather than a constant, a local server is not treated as unconfigured just because it has no API key, and its model list is read from the server, since only the server knows what you have pulled.
+
+### Extensions
+
+- **Decorations draw.** `createTextEditorDecorationType` used to return a working-looking handle that drew nothing — the last place in the shim that answered successfully and did nothing.
+- **`onDidChangeActiveTextEditor` handed over the wrong editor.** It fired before the editor existed and the notification was dropped on the way out, so opening a file delivered either nothing or the *previous* file's editor. Extensions that draw on editor changes were drawing onto whatever was open before.
+- **`WorkspaceEdit` can create, delete and rename files.** All-or-nothing: if one operation cannot be done, none are. A file with unsaved edits is neither deleted nor overwritten.
+- **`env.openExternal` and `env.clipboard` do their jobs** instead of reporting success and doing nothing.
+
+### Proposals in the code
+
+A pending proposal is marked on the line it changes; hovering shows why. Accept and reject are CodeLens on that line, not only buttons on the card. While an edit types itself in, a ghost caret shows the exact character being written.
+
+This completes the four editing-visualization pillars. The original plan for that phase was to fork Code-OSS; the fork never happened and should not, since this app has had its own renderer from the start. [docs/PHASE2-SURVEY.md](docs/PHASE2-SURVEY.md) audits the pillars against the code.
+
+### Also
+
+- The app survives a main-process crash instead of vanishing, and writes a crash log.
+- Async failures that used to disappear are surfaced.
+- The tour teaches both modes rather than only the one you are in.
+- Shell commands run only in the open workspace.
+- MCP resources and prompts are offered to the agent, not just listed.
+
+
 ## [0.2.0] — Everything the extension host said it could do
 
 The previous release fixed behaviour the app claimed to have. This one does the same thing one layer out: an extension could load, report success, and then do nothing at all, because most of what it asked for was answered by a function that returned an empty value.

README.md

@@ -93,13 +93,14 @@ same table the app dispatches from, so it cannot drift from what actually happen
 
 **Debugging** — Breakpoints, call stack, variables, and stepping via DAP (Python/debugpy today).
 
-**Extensions** — Install VS Code extensions from Open VSX, with TextMate grammars and icon themes. The `vscode` shim gives them the active editor for real: reading the document, moving the selection, and editing through the model so `Ctrl+Z` still works. Extensions can ask you things too — quick picks, input boxes with validation, and messages with buttons, each labelled with the extension doing the asking. Diagnostics land in the editor and the problems panel; definitions and formatters work; settings, file watchers and status bar items all do what they say. Trees and webviews contributed by an extension get their own place in the sidebar, webviews sandboxed. Schutz-native extensions can register commands, show panels, and subscribe to what happens in the IDE — files opened and saved, proposals accepted or rejected, agent turns starting and ending.
+**Extensions** — Install VS Code extensions from Open VSX, with TextMate grammars and icon themes. The `vscode` shim gives them the active editor for real: reading the document, moving the selection, and editing through the model so `Ctrl+Z` still works. Extensions can ask you things too — quick picks, input boxes with validation, and messages with buttons, each labelled with the extension doing the asking. Diagnostics land in the editor and the problems panel; definitions and formatters work; settings, file watchers and status bar items all do what they say. Trees and webviews contributed by an extension get their own place in the sidebar, webviews sandboxed. Quick fixes appear in the editor's own menu and apply through the model, so one `Ctrl+Z` takes them back. Schutz-native extensions can register commands, show panels, and subscribe to what happens in the IDE — files opened and saved, proposals accepted or rejected, agent turns starting and ending.
 
 **Localization** — The full UI ships in Korean, English, German, and Japanese.
 
 ## Design principles
 
-- **Provider-agnostic** — Claude, OpenAI, Grok, and GLM today, behind a swappable adapter
+- **Provider-agnostic** — Claude, OpenAI, Grok, GLM and Gemini today, plus any OpenAI-compatible
+  server you point it at, all behind a swappable adapter
 - **Observable by default** — every AI action surfaces in the UI
 - **Human-in-the-loop** — every change can be accepted, rejected, or reverted
 - **Progressive fidelity** — validate the experience first, then deepen it at the editor core
@@ -121,21 +122,44 @@ npm run dist:linux # Linux build
 ## Roadmap
 
 - **Phase 1** — validate the core experience *(done)*
-- **Phase 2** — deepen renderer-level visual effects at the editor core *(in progress)*
-- **Phase 3** — more providers (Gemini, local/OpenAI-compatible endpoints) and codebase indexing
+- **Phase 2** — the four editing-visualization pillars *(done)*
+- **Phase 3** — more providers *(done)* and codebase indexing *(next, as a 0.4 preview)*
+
+Phase 2 was originally written as "fork Code-OSS", on the assumption that the visual effects
+were impossible through the extension API. That fork never happened and should not: the app
+has its own renderer already (Electron + React + Monaco), so the reason for forking never
+applied. [docs/PHASE2-SURVEY.md](docs/PHASE2-SURVEY.md) audits the four pillars against the
+code and closes the fork item as not-applicable. The three things the design named and the
+code lacked — a ghost cursor, a rationale tooltip on the changed line, and accept/reject
+CodeLens — now exist.
+
+Providers: Claude, GPT, Grok, GLM, Gemini, and any OpenAI-compatible server you point it at —
+Ollama, LM Studio, llama.cpp, or a corporate proxy. A local server needs an address, not a key.
 
 The ecosystem work that used to sit in Phase 3 has shipped: VS Code extensions, Claude Code
 skills and subagents, MCP servers over stdio and HTTP, and `.mcpb` bundles are all in the app
 today. A VS Code extension can read and edit the active document, publish diagnostics, ask
 you a question, read its own settings, watch files, show status, and contribute a tree or
 webview to the sidebar. MCP speaks all three of its pillars rather than tools alone.
 
-Where it is still thin, specifically: code actions are not wired (the shim does not expose
-`WorkspaceEdit`, so an extension cannot build one), `TreeView.reveal` and `document.save()`
-are unimplemented, and a file that is not already in a tab cannot be opened. Each of those
-returns an error rather than an empty value, so an extension that needs one fails loudly
-instead of doing nothing. MCP resources and prompts are readable and shown in the panel, but
-are not yet offered to the agent as tools.
+MCP resources and prompts are offered to the agent, not just listed in a panel.
+
+Where it is still thin, specifically: a `WorkspaceEdit` cannot create, delete or rename
+files, and the shim has no custom editors and no debug adapter API. Those three fail loudly —
+`WorkspaceEdit` throws, and the other two are absent from the API object, so calling one is a
+`TypeError` rather than a quiet nothing.
+
+Decorations used to sit in that list as the one gap that answered successfully and drew
+nothing. They now draw: `createTextEditorDecorationType` compiles the requested styling into
+a real rule, `setDecorations` puts it on the Monaco editor, and passing an empty array clears
+what that type drew. Inline blame and coverage highlighting work, `before`/`after` content
+included.
+
+Fixing it surfaced the reason such an extension would have appeared broken anyway:
+`onDidChangeActiveTextEditor` fired before the editor existed, and the notification was
+dropped on the way out, so opening a file delivered either nothing or the *previous* file's
+editor. Extensions that draw on editor changes — which is most of them — were being handed
+the wrong editor. The event now waits for the pane and fires once per file.
 
 ## Contributing
 

docs/DESIGN.md

@@ -142,12 +142,22 @@ AI의 편집은 즉시 파일에 쓰이지 않고 **트랜잭션**으로 관리
 - 멀티파일 오버뷰 Webview (2.4)
 - **검증 질문**: 확장 API로 어디까지 "멋지게" 되는가? 무엇이 답답한가?
 
-### Phase 2 — 포크(Fork)로 이전
-> 목표: 확장으로 불가능했던 시각 효과를 위해 Code-OSS 포크. 진짜 차별화 완성.
-- Code-OSS / VSCodium 기반 브랜딩·빌드 파이프라인
-- 에디터 렌더러 레벨의 부드러운 타이핑 애니메이션 (2.1 완성)
-- 확장 코드를 코어 기능으로 통합
-- 배포(자동 업데이트, 서명 등)
+### Phase 2 — 자체 렌더러 위의 시각화 *(완료, 2026-08-05)*
+> 원래 계획은 "Code-OSS 포크" 였다. **하지 않았고, 하지 않는 것이 맞다.**
+> 확장 API 로 렌더러에 손댈 수 없다는 것이 포크의 이유였는데, 이 앱은
+> Electron + React + Monaco 로 처음부터 자기 렌더러를 갖고 있어 그 이유가 없다.
+> 판정과 근거는 [PHASE2-SURVEY.md](PHASE2-SURVEY.md).
+
+- ~~Code-OSS / VSCodium 기반 브랜딩·빌드 파이프라인~~ — 해당 없음(포크 안 함)
+- 타이핑 애니메이션 + 글로우·잔광 + **유령 커서** (2.1)
+- **변경 사유 툴팁**과 **수락/거절 CodeLens** — 바뀌는 그 줄에서 (2.2)
+- 작업 상태·계획 패널 (2.3)
+- 멀티파일 변경 개요 (2.4)
+- 배포: 설치본 5종(서명은 아직, 자동 업데이트는 의도적으로 안 씀)
+
+남은 것 하나 — 2.1 의 "스트리밍 삽입" 은 지금 **완성된 텍스트의 재생**이라 모델이
+실제로 쓰는 속도와 무관하다. 진짜 토큰 스트림을 흘리려면 제안·승인 흐름 자체를
+바꿔야 해서, 시각 효과가 아니라 별건으로 남긴다.
 
 ### Phase 3 — 확장·생태계
 - 멀티 프로바이더 완성
@@ -162,7 +172,7 @@ AI의 편집은 즉시 파일에 쓰이지 않고 **트랜잭션**으로 관리
 |---|---|---|
 | 언어 | TypeScript | VSCode 생태계 표준 |
 | Phase 1 베이스 | VSCode Extension API | 빠른 검증 |
-| Phase 2 베이스 | Code-OSS (VSCodium 방식) | 포크 |
+| Phase 2 베이스 | ~~Code-OSS (VSCodium 방식)~~ → Electron + React + Monaco | 포크하지 않음 — 렌더러가 이미 우리 것 |
 | UI (패널·오버뷰) | Webview + (React 등) | 커스텀 시각화 자유도 |
 | AI 통신 | 스트리밍 HTTP/SSE | 실시간성 필수 |
 | 상태 관리 | (미정) | 오케스트레이터 이벤트 기반 |
@@ -172,7 +182,7 @@ AI의 편집은 즉시 파일에 쓰이지 않고 **트랜잭션**으로 관리
 ## 6. 리스크 & 열린 질문
 
 **리스크**
-- **R1. 확장 API의 시각 표현 한계** — 2.1의 매끄러운 애니메이션이 확장에서 기대만큼 안 나올 수 있음. → Phase 1에서 조기 검증, 안 되면 Phase 2 앞당김.
+- **R1. 확장 API의 시각 표현 한계** — 2.1의 매끄러운 애니메이션이 확장에서 기대만큼 안 나올 수 있음. → Phase 1에서 조기 검증, 안 되면 Phase 2 앞당김. **(해소: 확장이 아니라 자체 렌더러로 갔다. 애니메이션·유령 커서·CodeLens 모두 우리 코드다.)**
 - **R2. 포크 유지보수 부담** — VSCode 업스트림 추종 비용. VSCodium 사례 참고로 최소화.
 - **R3. 실시간 스트리밍 편집의 충돌** — 사용자가 편집 중인데 AI도 편집하면? → 트랜잭션 모델 + 락/머지 정책 필요.