ManSio · Nov 11, 2026 · Nov 11, 2026 · Nov 11, 2026
diff --git a/AGENT_DIARY.md b/AGENT_DIARY.md
@@ -1757,3 +1757,11 @@ BREACH: MSC-P-19 | date: 2026-10-03 | rule: §12 | guard_at_time: no | detected_
 **Fix:** `_stderr_drain()` — фоновый поток с ring-buffer на 20 строк (пайп всегда вычитывается, хвост попадает в сообщение FAIL), плюс `LOG_LEVEL=WARNING` для пробы (`src/main.py:126` читает эту переменную) — сервер не заливает пайп на холодной машине. `_read_message` остался ограниченным по времени.
 **Проверка:** локально 5/5 PASS, 13 s; CI run `43092017919` (head `20f287b`) — **7/7 jobs success**: `test` (ubuntu + windows), `clean-state`, `manifest-parity`, `secret-scan`, `docker-build`, `transport-equiv-e2e`.
 **Урок:** дочерний процесс с `stdout/stderr=PIPE` обязан иметь читателя на каждом пайпе; «я прочитаю stderr только при ошибке» — это дедлок, который воспроизводится только там, где логов больше буфера.
+
+## [2026-11-12] G6: doctor/clean и text-first по умолчанию — «что стоит и сколько весит»
+**Симптом:** установщик предлагал ~657 МБ моделей при штатном `MSCODEBASE_DENSE_OFF=true`; команды «сколько это занимает и как убрать» не было; uninstaller существовал, но не был описан; лимиты кэшей жили в коде (`artifact_gc`, `log_manager`), а не в документации.
+**Fix:** `install.py --doctor [--json]` (слои: install root / data root / проектные артефакты, размеры через `os.walk`, число чанков из `chunks.db`, пути без создания каталогов) и `install.py --clean [--dry-run] [--yes] [--json]` с guard'ами (отказ на исходный чекаут, корень ФС, домашний каталог) и ненулевым exit при остатке (заблокированные файлы работающим хостом). Модели стали opt-in: `models_enabled()` + `--with-models`; `--skip-models/--no-models` остались совместимыми.
+**Замеры (реальная машина, 2026-11-12):** install root **2.7 GB / 21 372 файла**, data root **302.7 MB / 2 861 файл**, итого **3.0 GB / 24 233 файла**. `--sync-only` в пустой каталог: **871 файл / 18.7 MB за 3.1 с**.
+**Автотест DoD:** `tests/test_install_cli.py::test_real_sync_install_then_doctor_then_clean_leaves_zero` — реальный `--sync-only` → `--doctor --json` (install_root.files > 100) → `--clean --yes` → `remaining_bytes == 0`, каталог удалён. Синтетический кейс отдельно проверяет dry-run (ничего не удалено), отказ без `--yes` (exit 2) и `doctor` после clean (`total_bytes == 0`).
+**Найденный по ходу баг:** двойная обёртка stdout в UTF-8 (`_utf8_stdout()` в main + старый TUI-блок) закрывала буфер: `ValueError: I/O operation on closed file` на `--sync-only`. Убран дубль, оставлена одна обёртка на все режимы.
+**Урок:** «очистка по слоям» без dry-run и guard'ов — это удаление чужого кэша по недопониманию; ноль остатка надо доказывать на файловой системе (`remaining_bytes == 0`), а не на числе вызванных `rmtree`.
diff --git a/README.md b/README.md
@@ -231,14 +231,22 @@ MCP host: OpenCode (primary) or Zed
   └── Optional      embedding providers, reranker, LSP rename
 ```
 
-Canonical data locations (outside the repository):
+Canonical data locations (four layers; full inventory with sizes and policies in
+[docs/en/DATA_MAP.md](docs/en/DATA_MAP.md)):
 
-- `%LOCALAPPDATA%/mscodebase/projects/<hash>/` — per-project graph, memory,
+- install root (`MSCODEBASE_INSTALL_ROOT`, default the historical Zed extension
+  directory) — the code copy, `venv/`, optional ONNX/GGUF models, uninstaller.
+- `%LOCALAPPDATA%/mscodebase/projects/<hash8>/` — per-project graph, memory,
   job state, commit memory, telemetry.
-- `%LOCALAPPDATA%/mscodebase/logs/` — server logs.
+- `%LOCALAPPDATA%/mscodebase/logs/` — server logs (2 MB × 3 rotation, 7-day retention).
+- `%LOCALAPPDATA%/mscodebase/models/`, `repos/` — shared model cache and LRU(5)+24 h
+  repo clones.
 - `<project>/.mscodebase/chunks.db` — the text chunk store used by the default
   search path.
 
+`python install.py --doctor` prints every layer with its measured size;
+`python install.py --clean --dry-run` lists what `--clean` would delete.
+
 LanceDB is part of the optional vector path and is not required by the default
 text-first configuration.
 
@@ -343,7 +351,8 @@ stale/empty index, and gold-label mismatch.
 | [docs/FACTS.md](docs/FACTS.md) | Every public number and the command that regenerates it |
 | [docs/en/ARCHITECTURE.md](docs/en/ARCHITECTURE.md) | Layers, DI, module boundaries, tool profile |
 | [docs/en/SEARCH_PIPELINE.md](docs/en/SEARCH_PIPELINE.md) | Retrieval, ranking, docs routing, investigate fan-out |
-| [docs/en/INSTALL.md](docs/en/INSTALL.md) | Installation, the sync/restart loop, troubleshooting |
+| [docs/en/INSTALL.md](docs/en/INSTALL.md) | Installation, the sync/restart loop, doctor/clean, troubleshooting |
+| [docs/en/DATA_MAP.md](docs/en/DATA_MAP.md) | Every path the installer/server creates, with sizes and retention policies |
 | [docs/en/CONFIGURATION.md](docs/en/CONFIGURATION.md) | Every environment variable and its default |
 | [docs/en/CLIENTS.md](docs/en/CLIENTS.md) | OpenCode, Zed, Claude Code, VS Code, CLI, Docker |
 | [docs/en/HOOKS_AND_GATES.md](docs/en/HOOKS_AND_GATES.md) | The OpenCode plugin, the 12 pre-commit gates, CI jobs |

diff --git a/docs/en/CLIENTS.md b/docs/en/CLIENTS.md
@@ -35,7 +35,7 @@ Linux   : ~/.local/share/zed/extensions/mscodebase-intelligence
 
 Move it anywhere with `MSCODEBASE_INSTALL_ROOT=<path>`; caches, logs and per-project
 artifacts live elsewhere (`MSCODEBASE_DATA_DIR`, default `%LOCALAPPDATA%\mscodebase`).
-
+The full inventory is in [DATA_MAP.md](DATA_MAP.md).
 
 Alternative without the installer (any Python package manager that exposes console
 scripts):

diff --git a/docs/en/DATA_MAP.md b/docs/en/DATA_MAP.md
@@ -0,0 +1,111 @@
+# Where MSCodeBase puts data
+
+> Written 2026-11-12 and verified against `install.py` v3.2,
+> `src/core/artifact_paths.py`, `src/core/artifact_gc.py`,
+> `src/core/log_manager.py`, `src/sources/git_url/__init__.py` and a live
+> `install.py --doctor` run. Numbers are measured, not estimated.
+
+One question this page answers: *what did the installer and the server create on
+my disk, who owns each path, and how do I get rid of it?*
+
+```bash
+python install.py --doctor              # what is installed and how much it weighs
+python install.py --clean --dry-run     # paths and sizes; deletes nothing
+python install.py --clean --yes         # delete every layer below
+```
+
+## 1. The four layers
+
+| Layer | Path (default) | Created by | What it is |
+|---|---|---|---|
+| **Install root** | Windows `%LOCALAPPDATA%\Zed\extensions\mscodebase-intelligence`; macOS `~/Library/Application Support/Zed/extensions/…`; Linux `~/.local/share/zed/extensions/…` | `install.py` steps `step_copy`, `step_venv`, `step_pip`, `step_llama`, `step_gguf`, `step_models`, `step_uninst` | The installed copy: code, `venv/`, optional models, uninstaller |
+| **Data root** | `%LOCALAPPDATA%\mscodebase` (Win) · `$XDG_CACHE_HOME/mscodebase` or `~/.cache/mscodebase` (POSIX) | `src/core/artifact_paths.py::get_data_root` | Per-project caches, logs, shared models, repo clones |
+| **Project artifacts** | `<project>/.mscodebase/chunks.db` | server (`MSCODEBASE_SQLITE_CHUNKS`) | FTS5 text index for that project |
+| **Host config** | Zed: `%APPDATA%\Zed\settings.json` (`context_servers`) | `install.py` step `step_zedcfg` → `adapters/zed/zed_config.py` | One entry telling the host how to launch the server |
+
+Both roots are relocatable: `MSCODEBASE_INSTALL_ROOT` and `MSCODEBASE_DATA_DIR`
+(`BASE_INDEX_DIR` is still accepted as a legacy alias for the second).
+
+## 2. Install root — what is inside
+
+| Path | Owner | Purpose | Size |
+|---|---|---|---|
+| `src/`, `scripts/`, `adapters/`, `locales/`, `docs/` | `step_copy` (`_sync_dir`) | Executable copy of the repository (without `experiments/`) | a few MB |
+| `venv/` | `step_venv`, `step_pip` | Python environment with pinned dependencies | ~1 GB |
+| `venv/Scripts/mscodebase-mcp.exe` | `pip` entry point (`pyproject.toml [project.scripts]`) | The stdio command hosts launch | — |
+| `.codebase_models/onnx/…` | `step_models` (only with `--with-models`) | ONNX embedder + reranker | ~657 MB |
+| `llama_msvc/` | `step_llama` | llama-server binary for the GGUF path | ~200 MB |
+| `models/*.gguf` | `step_gguf` (only with `--with-models`) | GGUF embedder + reranker | ~700 MB |
+| `uninstall.bat` / `uninstall.sh` | `step_uninst` (`_build_uninstaller`) | Kills the server and removes this directory | < 1 KB |
+| `__mscodebase_ext__.marker` | `step_copy` | Tells `src/main.py` it runs from an installed copy | < 1 KB |
+
+Measured on the development machine (2026-11-12, models installed):
+**3.0 GB over 28 429 files**, of which `venv/` plus models make up almost all of it.
+
+## 3. Data root — what is inside
+
+| Path | Owner | Purpose | Policy |
+|---|---|---|---|
+| `projects/<hash8>/graph.db` | `artifact_paths.get_graph_db_path` | PropertyGraph: files, symbols, calls, imports | deleted with the project dir |
+| `projects/<hash8>/lancedb_v2/` | `get_index_dir`, `get_db_path` | Vector index (only when dense is enabled) | same |
+| `projects/<hash8>/intelligence/` | `get_intelligence_dir` | Project memory: ADRs, incidents, debt | same |
+| `projects/<hash8>/metrics/` | `get_metrics_dir` | Indexing job history (ETA model) | same |
+| `projects/<hash8>/commit_memory/` | `get_commit_memory_dir` | Semantic commit cache | same |
+| `projects/<hash8>/branches/` | `get_branches_dir` | Per-branch index state | same |
+| `projects/<hash8>/telemetry/` | `get_telemetry_dir` | Daily JSON snapshots | files older than **90 days** pruned |
+| `projects/<hash8>/summaries_cache/` | `get_summaries_cache_dir` | LLM chunk summaries | same as project dir |
+| `projects/<hash8>/progress.json` | `get_progress_file` | Indexing progress contract | overwritten |
+| `projects/<hash8>/action_receipts.jsonl` | `src/core/action_receipt.py` | Receipts for write actions | verified 60 d / inconclusive 7 d (`purge`) |
+| `logs/mscodebase-intelligence.log` | `log_manager.get_main_log_path` | Server log | rotates at 2 MB × 3 files, deleted after **7 days** |
+| `logs/crash.json` | `get_crash_log_path` | Last resource-monitor crash | overwritten |
+| `models/.codebase_models/onnx/…` | `get_onnx_models_base` | Shared copy of downloaded models | only `--clean` |
+| `repos/<hash8>/` | `get_repos_cache_dir` | Clones of remote repositories | LRU(5) + TTL 24 h (`src/sources/git_url`) |
+| `projects/<hash8>/` (whole dir) | — | — | inactive dirs older than **30 days** and empty dirs are pruned by `artifact_gc.prune_stale_artifacts()` |
+
+Measured on the development machine (2026-11-12): **302.9 MB over 2 934 files**.
+`--doctor` in a real project directory that also holds `chunks.db` (78 MB) and a
+local model copy (960 MB) reported **4.3 GB over 31 385 files** across all layers.
+
+## 4. Project artifacts
+
+| Path | Owner | Purpose |
+|---|---|---|
+| `<project>/.mscodebase/chunks.db` | `MSCODEBASE_SQLITE_CHUNKS` storage | Text index (FTS5/BM25) used by the default search path; `--doctor` reports its chunk count |
+| `<project>/.codebase_indices/`, `.codebase/`, `.mscodebase/telemetry/` | legacy | Moved into the data root on first project open (`artifact_paths.migrate_legacy_artifacts`) |
+| `<project>/.codebase_models/` | `scripts/download_model.py` (download staging) | Temporary copy before the model is installed under the install root |
+
+`experiments/**` is excluded from indexing by design, so it never grows the index.
+
+## 5. Policies
+
+- **Models are opt-in.** `install.py` is text-first: without `--with-models` it
+  performs no model download and writes no model directory. `--skip-models` and
+  `--no-models` are still accepted for old scripts.
+- **Logs** rotate by size (2 MB, 3 files) and are pruned after 7 days
+  (`log_manager._MAX_LOG_BYTES`, `_MAX_LOG_FILES`, `_LOG_RETENTION_DAYS`).
+- **Telemetry** snapshots are pruned after 90 days; **inactive project dirs**
+  after 30 days (`artifact_gc.DEFAULT_*`); empty project dirs are removed at once.
+- **Repo clones** are an LRU of 5 with a 24 h TTL, plus a 500 MB post-clone size
+  cap (`git_url.DEFAULT_MAX_CLONE_BYTES`).
+- **Shared model cache** (`data_root/models`) has no TTL: models are removed only
+  by `install.py --clean`.
+- **Uninstall** is `install.py --clean --yes`, which removes all four layers
+  except the host config entry (delete the `mscodebase` entry from
+  `settings.json` / `opencode.json` / `.mcp.json` yourself, or run
+  `install.py --no-zed` and never create it).
+
+## 6. Reading the doctor output
+
+```
++ install_root              3.0 GB  28429 file(s)  C:\Users\<user>\AppData\Local\Zed\extensions\mscodebase-intelligence
++ data_root               302.9 MB   2934 file(s)  C:\Users\<user>\AppData\Local\mscodebase
++ project_chunks           78.0 MB      4 file(s)  D:\work\repo\.mscodebase
+   ...
+
+index           7278 chunk(s)  (D:\work\repo\.mscodebase\chunks.db)
+total           4.3 GB over 31385 file(s)
+```
+
+`+` exists, `-` absent. `index missing` means the text index for the current
+directory was never built (the first `codebase(action="index")` call builds it),
+not that search is broken. `--json` prints the same data for scripts.