feat(distill): add OpenAI-compatible local LLM backend

CLAUDE.md

@@ -90,10 +90,22 @@ query → BM25 (FTS5 verbatim) + HNSW (distilled embedding)
 
 ```toml
 [distill]
-model = "claude-haiku-4-5-20251001"   # 蒸留に使うモデル
+provider = "claude"                   # 蒸留 backend: "claude" | "openai"（既定 "claude"）
+model = "claude-haiku-4-5-20251001"   # 蒸留に使うモデル（provider 共通）
 batch_limit = 20                       # hook 1回あたりの蒸留上限
 ```
 
+ローカル LLM（OpenAI 互換: Ollama / LM Studio / llama.cpp-server / vLLM）で蒸留する場合は `provider = "openai"` と `base_url` を指定する（新規依存なし・API キー不要＝`Authorization` ヘッダーは送らないローカル専用）:
+
+```toml
+[distill]
+provider = "openai"
+model = "qwen2.5:7b"
+base_url = "http://localhost:11434/v1"   # Ollama（LM Studio は :1234/v1）
+```
+
+`provider = "openai"` のとき `base_url` は必須。未設定/空なら警告して `claude` にフォールバックする。`provider = "claude"`（既定）では `base_url` は無視され、従来どおり `claude --print` で蒸留する。実装は `llm.py` の `DistillBackend` + `call_claude` dispatcher（`_call_openai` は stdlib `urllib`、`response_format=json_object` + プロンプト内スキーマ + `_validate_palace` 検証 + 1 回リトライ）。
+
 ## CLI Commands
 
 ```bash

README.md

@@ -145,6 +145,7 @@ After `loci init` (or `loci hook install`), everything runs automatically:
 
 ```toml
 [distill]
+provider = "claude"                    # Distillation backend: "claude" | "openai" (default "claude")
 model = "claude-haiku-4-5"             # Model for distillation (default)
 batch_limit = 20                       # Max distillations per hook run
 min_chars = 100                        # Skip distillation for exchanges shorter than this
@@ -155,6 +156,20 @@ min_chars = 50                         # Skip indexing exchanges shorter than th
 
 There are two `min_chars` settings: `[index] min_chars` controls what gets indexed at all, while `[distill] min_chars` further skips distillation (the LLM cost) for short exchanges that were already indexed.
 
+### Distilling with a local LLM
+
+Distillation is a small per-exchange structured-extraction task, so a local model is usually good enough. Any OpenAI-compatible endpoint (Ollama, LM Studio, llama.cpp-server, vLLM) works by setting `provider = "openai"` and `base_url` — no new dependencies, no API key (the `Authorization` header is never sent, so this is local-only):
+
+```toml
+[distill]
+provider = "openai"
+model = "qwen2.5:7b"
+base_url = "http://localhost:11434/v1"   # Ollama
+# base_url = "http://localhost:1234/v1"  # LM Studio
+```
+
+`base_url` is required when `provider = "openai"`; if it is missing or empty the provider falls back to `claude` with a warning. With `provider = "claude"` (the default), `base_url` is ignored and distillation runs through `claude --print` as before.
+
 ## Acknowledgments
 
 The palace object model, room-based topic grouping, and BM25+HNSW fusion search are based on:

src/codeatrium/cli/__init__.py

@@ -130,9 +130,16 @@ def init(
                 "# Codeatrium configuration\n"
                 "\n"
                 "[distill]\n"
+                '# provider = "claude"   # 蒸留 LLM: "claude" | "openai"（既定 "claude"）\n'
                 '# model = "claude-haiku-4-5-20251001"\n'
                 "# batch_limit = 20\n"
                 "# min_chars = 100   # この文字数未満の exchange は蒸留スキップ\n"
+                "#\n"
+                "# ローカル LLM（OpenAI 互換）で蒸留する場合:\n"
+                '#   provider = "openai"\n'
+                '#   model = "qwen2.5:7b"\n'
+                '#   base_url = "http://localhost:11434/v1"   # Ollama\n'
+                '#   # base_url = "http://localhost:1234/v1"  # LM Studio\n'
                 "\n"
                 "[index]\n"
                 "# min_chars = 50   # trivial フィルタ閾値（文字数）\n"
@@ -225,6 +232,7 @@ def init(
         try:
             from codeatrium.config import load_config
             from codeatrium.distiller import distill_all
+            from codeatrium.llm import DistillBackend
 
             cfg = load_config(root)
             typer.echo("Running distillation...")
@@ -239,7 +247,7 @@ def _on_progress(
 
             count, err_count = distill_all(
                 db,
-                model=cfg.distill_model,
+                backend=DistillBackend.from_config(cfg),
                 on_progress=_on_progress,
                 project_root=str(root),
                 distill_min_chars=cfg.distill_min_chars,

src/codeatrium/cli/distill_cmd.py

@@ -1,4 +1,4 @@
-"""loci distill コマンド"""
+"""loci distill コマンド — backend/provider サポート"""
 
 from __future__ import annotations
 
@@ -19,6 +19,7 @@ def distill(
 
     from codeatrium.config import load_config
     from codeatrium.distiller import distill_all
+    from codeatrium.llm import DistillBackend
     from codeatrium.paths import db_path, find_project_root
 
     root = find_project_root()
@@ -59,7 +60,7 @@ def _on_progress(cur: int, tot: int, error: str | None = None) -> None:
         count, err_count = distill_all(
             db,
             limit=limit,
-            model=cfg.distill_model,
+            backend=DistillBackend.from_config(cfg),
             on_progress=_on_progress,
             project_root=str(root),
             distill_min_chars=cfg.distill_min_chars,

src/codeatrium/config.py

@@ -1,4 +1,6 @@
-"""設定ファイルの読み込み — .codeatrium/config.toml"""
+"""設定ファイルの読み込み — .codeatrium/config.toml
+
+Supports distill_provider (claude/openai) and distill_base_url for provider switching."""
 
 from __future__ import annotations
 
@@ -16,6 +18,8 @@
 DEFAULT_DISTILL_BATCH_LIMIT = 20
 DEFAULT_INDEX_MIN_CHARS = 50
 DEFAULT_DISTILL_MIN_CHARS = 100
+DEFAULT_DISTILL_PROVIDER = "claude"
+VALID_DISTILL_PROVIDERS = frozenset({"claude", "openai"})
 
 
 @dataclass
@@ -26,6 +30,8 @@ class Config:
     distill_batch_limit: int = DEFAULT_DISTILL_BATCH_LIMIT
     index_min_chars: int = DEFAULT_INDEX_MIN_CHARS
     distill_min_chars: int = DEFAULT_DISTILL_MIN_CHARS
+    distill_provider: str = "claude"
+    distill_base_url: str | None = None
 
 
 def load_config(project_root: Path) -> Config:
@@ -79,9 +85,31 @@ def load_config(project_root: Path) -> Config:
         )
         distill_min_chars = DEFAULT_DISTILL_MIN_CHARS
 
+    provider = distill.get("provider", DEFAULT_DISTILL_PROVIDER)
+    if not isinstance(provider, str) or provider not in VALID_DISTILL_PROVIDERS:
+        print(
+            "Warning: distill.provider must be one of {claude, openai}, using default.",
+            file=sys.stderr,
+        )
+        provider = DEFAULT_DISTILL_PROVIDER
+
+    base_url = distill.get("base_url")
+    if base_url is not None and not isinstance(base_url, str):
+        base_url = None
+
+    if provider == "openai" and (base_url is None or not base_url.strip()):
+        print(
+            "Warning: distill.provider is 'openai' but base_url is not set, falling back to 'claude'.",
+            file=sys.stderr,
+        )
+        provider = DEFAULT_DISTILL_PROVIDER
+        base_url = None
+
     return Config(
         distill_model=model,
         distill_batch_limit=batch_limit,
         index_min_chars=min_chars,
         distill_min_chars=distill_min_chars,
+        distill_provider=provider,
+        distill_base_url=base_url,
     )

src/codeatrium/distiller.py

@@ -5,6 +5,8 @@
   ② claude -p で palace object 生成（--output-format json --json-schema）
   ③ distill_text を embedding して vec_palace に登録
   ④ files_touched を tree-sitter で解析してシンボルを symbols テーブルに登録
+
+backend パラメータで蒸留に使う DistillBackend を指定可能。
 """
 
 from __future__ import annotations
@@ -19,7 +21,7 @@
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
 
 from codeatrium.embedder import Embedder, EmbedderSetupError
-from codeatrium.llm import DISTILL_PROMPT_TEMPLATE, call_claude
+from codeatrium.llm import DISTILL_PROMPT_TEMPLATE, DistillBackend, call_claude
 from codeatrium.models import PalaceObject
 from codeatrium.utils import sha256
 
@@ -92,7 +94,7 @@ def distill_exchange(
     agent_content: str,
     ply_start: int,
     ply_end: int,
-    model: str | None = None,
+    backend: DistillBackend | None = None,
     project_root: str | None = None,
 ) -> PalaceObject:
     """1つの exchange を蒸留して PalaceObject を返す
@@ -104,7 +106,7 @@ def distill_exchange(
         agent_content: エージェントコンテンツ
         ply_start: 開始ply
         ply_end: 終了ply
-        model: 蒸留に使うモデル（デフォルトはconfig.toml から）
+        backend: 蒸留に使う DistillBackend（省略時はデフォルト claude）
         project_root: プロジェクトルート（ファイルパスフィルタ用）
     """
     from codeatrium.db import get_connection
@@ -115,7 +117,7 @@ def distill_exchange(
         ply_end=ply_end,
         messages_text=messages_text,
     )
-    raw = call_claude(prompt, model=model)
+    raw = call_claude(prompt, backend=backend)
 
     # PRIMARY: exchange_files から読み込み
     con = get_connection(db_path)
@@ -302,13 +304,14 @@ def save_palace_object(
 def distill_all(
     db_path: Path,
     limit: int | None = None,
-    model: str | None = None,
+    backend: DistillBackend | None = None,
     on_progress: Callable[..., None] | None = None,
     project_root: str | None = None,
     distill_min_chars: int = 100,
 ) -> tuple[int, int]:
     """未蒸留の exchange を処理する。
 
+    backend: 蒸留に使う DistillBackend（省略時はデフォルト claude）
     distill_min_chars: この文字数未満の exchange は蒸留スキップ（デフォルト100）
     on_progress: (current, total, error=None) を受け取るコールバック
     Returns: (処理した exchange 数, エラー数)
@@ -360,7 +363,7 @@ def distill_all(
                 row["agent_content"],
                 row["ply_start"],
                 row["ply_end"],
-                model=model,
+                backend=backend,
                 project_root=project_root,
             )
             distill_text = palace.exchange_core + "\n" + palace.specific_context