kiri-mcp-server 0.24.4 → 0.25.0

package/dist/src/server/handlers.js

@@ -589,6 +589,7 @@ const HINT_SUBSTRING_BOOST = serverConfig.hints.substring.boost;
 const HINT_LOG_ENABLED = process.env.KIRI_HINT_LOG === "1";
 const HINT_DICTIONARY_ENABLED = process.env.KIRI_HINT_DICTIONARY !== "0";
 const HINT_DICTIONARY_LIMIT = Math.max(0, Number.parseInt(process.env.KIRI_HINT_DICTIONARY_LIMIT ?? "2", 10));
+const DEFAULT_FILECACHE_MAX_BYTES = 64 * 1024 * 1024;
 // Issue #68: Path/Large File Penalty configuration (環境変数で上書き可能)
 const PATH_MISS_DELTA = serverConfig.penalties.pathMissDelta;
 const LARGE_FILE_DELTA = serverConfig.penalties.largeFileDelta;
@@ -1325,6 +1326,16 @@ function parseEmbedding(vectorJson, vectorDims) {
         return null;
     }
 }
+function coerceLineCount(value) {
+    if (value === null || value === undefined) {
+        return null;
+    }
+    const numeric = typeof value === "bigint" ? Number(value) : value;
+    if (!Number.isFinite(numeric) || numeric < 0) {
+        return null;
+    }
+    return numeric;
+}
 function applyStructuralScores(candidates, queryEmbedding, structuralWeight) {
     if (!queryEmbedding || structuralWeight <= 0) {
         return;
@@ -1501,7 +1512,7 @@ async function fetchEmbeddingMap(db, repoId, paths) {
 }
 async function loadFileContent(db, repoId, filePath) {
     const rows = await db.all(`
-      SELECT f.path, f.lang, f.ext, f.is_binary, b.content, fe.vector_json, fe.dims AS vector_dims
+      SELECT f.path, f.lang, f.ext, f.is_binary, b.content, b.line_count, fe.vector_json, fe.dims AS vector_dims
       FROM file f
       JOIN blob b ON b.hash = f.blob_hash
       LEFT JOIN file_embedding fe
@@ -1514,7 +1525,8 @@ async function loadFileContent(db, repoId, filePath) {
     if (!row || row.is_binary || row.content === null) {
         return null;
     }
-    const totalLines = row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length;
+    const totalLines = coerceLineCount(row.line_count) ??
+        (row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length);
     return {
         content: row.content,
         lang: row.lang,
@@ -1523,6 +1535,28 @@ async function loadFileContent(db, repoId, filePath) {
         embedding: parseEmbedding(row.vector_json ?? null, row.vector_dims ?? null),
     };
 }
+async function loadFileMetadata(db, repoId, filePath) {
+    const rows = await db.all(`
+      SELECT f.path, f.lang, f.ext, f.is_binary, b.line_count, fe.vector_json, fe.dims AS vector_dims
+      FROM file f
+      JOIN blob b ON b.hash = f.blob_hash
+      LEFT JOIN file_embedding fe
+        ON fe.repo_id = f.repo_id
+       AND fe.path = f.path
+      WHERE f.repo_id = ? AND f.path = ?
+      LIMIT 1
+    `, [repoId, filePath]);
+    const row = rows[0];
+    if (!row || row.is_binary) {
+        return null;
+    }
+    return {
+        lang: row.lang,
+        ext: row.ext,
+        totalLines: coerceLineCount(row.line_count),
+        embedding: parseEmbedding(row.vector_json ?? null, row.vector_dims ?? null),
+    };
+}
 function selectSnippet(snippets, matchLine) {
     const firstSnippet = snippets[0];
     if (!firstSnippet) {
@@ -1814,7 +1848,7 @@ async function fetchMetadataOnlyCandidates(db, tableAvailability, repoId, filter
         params.push(...clause.params);
     }
     const sql = `
-    SELECT f.path, f.lang, f.ext, b.content
+    SELECT f.path, f.lang, f.ext, b.content, b.line_count
     FROM file f
     JOIN blob b ON b.hash = f.blob_hash
     WHERE ${whereClauses.join(" AND ")}
@@ -2074,7 +2108,7 @@ async function fetchPathFallbackCandidates(db, repoId, terms, limit) {
     const filters = terms.map(() => "f.path ILIKE ?").join(" OR ");
     const params = [repoId, ...terms.map((term) => `%${term}%`), limit];
     return await db.all(`
-      SELECT f.path, f.lang, f.ext, f.is_binary, b.content, fe.vector_json, fe.dims AS vector_dims
+      SELECT f.path, f.lang, f.ext, f.is_binary, b.content, b.line_count, fe.vector_json, fe.dims AS vector_dims
       FROM file f
       JOIN blob b ON b.hash = f.blob_hash
       LEFT JOIN file_embedding fe
@@ -2607,6 +2641,40 @@ function readPenaltyFlags() {
         largeFilePenalty: process.env.KIRI_LARGE_FILE_PENALTY === "1",
     };
 }
+function parseByteSize(raw) {
+    const trimmed = raw.trim().toLowerCase();
+    const match = /^(\d+)([kmg]?)(b)?$/.exec(trimmed);
+    if (!match) {
+        return null;
+    }
+    const value = Number.parseInt(match[1] ?? "", 10);
+    if (!Number.isFinite(value)) {
+        return null;
+    }
+    const unit = match[2] ?? "";
+    let multiplier = 1;
+    if (unit === "k") {
+        multiplier = 1024;
+    }
+    else if (unit === "m") {
+        multiplier = 1024 * 1024;
+    }
+    else if (unit === "g") {
+        multiplier = 1024 * 1024 * 1024;
+    }
+    return value * multiplier;
+}
+function readFileCacheMaxBytes() {
+    const raw = process.env.KIRI_FILECACHE_MAX_BYTES;
+    if (raw === undefined) {
+        return DEFAULT_FILECACHE_MAX_BYTES;
+    }
+    const parsed = parseByteSize(raw);
+    if (parsed === null || Number.isNaN(parsed)) {
+        return DEFAULT_FILECACHE_MAX_BYTES;
+    }
+    return Math.max(0, parsed);
+}
 /**
  * クエリ統計を計算（単語数と平均単語長）
  */
@@ -2970,6 +3038,67 @@ async function contextBundleImpl(context, params) {
     const candidates = new Map();
     const stringMatchSeeds = new Set();
     const fileCache = new Map();
+    const fileCacheMaxBytes = readFileCacheMaxBytes();
+    let fileCacheBytes = 0;
+    const estimateContentBytes = (content) => Buffer.byteLength(content, "utf8");
+    const touchFileCache = (filePath) => {
+        const entry = fileCache.get(filePath);
+        if (!entry) {
+            return undefined;
+        }
+        fileCache.delete(filePath);
+        fileCache.set(filePath, entry);
+        return entry;
+    };
+    const pruneFileCache = () => {
+        if (fileCacheMaxBytes <= 0) {
+            if (fileCacheBytes <= 0) {
+                return;
+            }
+            for (const entry of fileCache.values()) {
+                if (!entry.contentBytes) {
+                    continue;
+                }
+                delete entry.content;
+                delete entry.contentBytes;
+            }
+            fileCacheBytes = 0;
+            return;
+        }
+        if (fileCacheBytes <= fileCacheMaxBytes) {
+            return;
+        }
+        for (const entry of fileCache.values()) {
+            if (!entry.contentBytes) {
+                continue;
+            }
+            fileCacheBytes = Math.max(0, fileCacheBytes - entry.contentBytes);
+            delete entry.content;
+            delete entry.contentBytes;
+            if (fileCacheBytes <= fileCacheMaxBytes) {
+                break;
+            }
+        }
+    };
+    const setFileCacheEntry = (filePath, next) => {
+        const existing = fileCache.get(filePath);
+        if (existing?.contentBytes) {
+            fileCacheBytes = Math.max(0, fileCacheBytes - existing.contentBytes);
+        }
+        const merged = { ...existing, ...next };
+        if (fileCacheMaxBytes <= 0 || !merged.content) {
+            delete merged.content;
+            delete merged.contentBytes;
+        }
+        else {
+            const contentBytes = estimateContentBytes(merged.content);
+            merged.contentBytes = contentBytes;
+            fileCacheBytes += contentBytes;
+        }
+        fileCache.delete(filePath);
+        fileCache.set(filePath, merged);
+        pruneFileCache();
+    };
     // Phase 2: IDF重み付けプロバイダーの初期化
     // キーワードの文書頻度に基づいて重みを計算し、高頻度語を自動的に減衰
     const idfProvider = createIdfProvider(db, repoId);
@@ -3006,7 +3135,7 @@ async function contextBundleImpl(context, params) {
             phraseParams.push(`${pathPrefix}%`);
         }
         const rows = await db.all(`
-        SELECT f.path, f.lang, f.ext, f.is_binary, b.content, fe.vector_json, fe.dims AS vector_dims
+        SELECT f.path, f.lang, f.ext, f.is_binary, b.content, b.line_count, fe.vector_json, fe.dims AS vector_dims
         FROM file f
         JOIN blob b ON b.hash = f.blob_hash
         LEFT JOIN file_embedding fe
@@ -3054,18 +3183,18 @@ async function contextBundleImpl(context, params) {
             const { line } = buildPreview(row.content, matchedPhrases[0]);
             candidate.matchLine =
                 candidate.matchLine === null ? line : Math.min(candidate.matchLine, line);
-            candidate.content ??= row.content;
             candidate.lang ??= row.lang;
             candidate.ext ??= row.ext;
-            candidate.totalLines ??= row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length;
+            const totalLines = coerceLineCount(row.line_count) ??
+                (row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length);
+            candidate.totalLines ??= totalLines;
             candidate.embedding ??= parseEmbedding(row.vector_json ?? null, row.vector_dims ?? null);
             stringMatchSeeds.add(row.path);
             if (!fileCache.has(row.path)) {
-                fileCache.set(row.path, {
-                    content: row.content,
+                setFileCacheEntry(row.path, {
                     lang: row.lang,
                     ext: row.ext,
-                    totalLines: candidate.totalLines ?? 0,
+                    totalLines: candidate.totalLines,
                     embedding: candidate.embedding,
                 });
             }
@@ -3088,7 +3217,7 @@ async function contextBundleImpl(context, params) {
             keywordParams.push(`${pathPrefix}%`);
         }
         const rows = await db.all(`
-        SELECT f.path, f.lang, f.ext, f.is_binary, b.content, fe.vector_json, fe.dims AS vector_dims
+        SELECT f.path, f.lang, f.ext, f.is_binary, b.content, b.line_count, fe.vector_json, fe.dims AS vector_dims
         FROM file f
         JOIN blob b ON b.hash = f.blob_hash
         LEFT JOIN file_embedding fe
@@ -3131,18 +3260,18 @@ async function contextBundleImpl(context, params) {
             const { line } = buildPreview(row.content, matchedKeywords[0]);
             candidate.matchLine =
                 candidate.matchLine === null ? line : Math.min(candidate.matchLine, line);
-            candidate.content ??= row.content;
             candidate.lang ??= row.lang;
             candidate.ext ??= row.ext;
-            candidate.totalLines ??= row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length;
+            const totalLines = coerceLineCount(row.line_count) ??
+                (row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length);
+            candidate.totalLines ??= totalLines;
             candidate.embedding ??= parseEmbedding(row.vector_json ?? null, row.vector_dims ?? null);
             stringMatchSeeds.add(row.path);
             if (!fileCache.has(row.path)) {
-                fileCache.set(row.path, {
-                    content: row.content,
+                setFileCacheEntry(row.path, {
                     lang: row.lang,
                     ext: row.ext,
-                    totalLines: candidate.totalLines ?? 0,
+                    totalLines: candidate.totalLines,
                     embedding: candidate.embedding,
                 });
             }
@@ -3185,19 +3314,21 @@ async function contextBundleImpl(context, params) {
             candidate.matchLine ??= 1;
             candidate.lang ??= row.lang;
             candidate.ext ??= row.ext;
-            candidate.totalLines ??= row.content?.split(/\r?\n/).length ?? null;
-            candidate.content ??= row.content;
+            const totalLines = coerceLineCount(row.line_count) ??
+                (row.content ? (row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length) : null);
+            if (totalLines !== null) {
+                candidate.totalLines ??= totalLines;
+            }
             candidate.embedding ??= parseEmbedding(row.vector_json ?? null, row.vector_dims ?? null);
             if (boostProfile !== "none") {
                 applyBoostProfile(candidate, row, profileConfig, weights, extractedTerms);
             }
             stringMatchSeeds.add(row.path);
-            if (!fileCache.has(row.path) && row.content) {
-                fileCache.set(row.path, {
-                    content: row.content,
+            if (!fileCache.has(row.path)) {
+                setFileCacheEntry(row.path, {
                     lang: row.lang,
                     ext: row.ext,
-                    totalLines: candidate.totalLines ?? 0,
+                    totalLines: candidate.totalLines,
                     embedding: candidate.embedding,
                 });
             }
@@ -3355,27 +3486,31 @@ async function contextBundleImpl(context, params) {
             if (isSuppressedPath(candidate.path)) {
                 continue;
             }
-            if (!candidate.content) {
-                const cached = fileCache.get(candidate.path);
-                if (cached) {
-                    candidate.content = cached.content;
-                    candidate.lang = cached.lang;
-                    candidate.ext = cached.ext;
-                    candidate.totalLines = cached.totalLines;
-                    candidate.embedding = cached.embedding;
-                }
-                else {
-                    const loaded = await loadFileContent(db, repoId, candidate.path);
-                    if (!loaded) {
-                        continue;
-                    }
-                    candidate.content = loaded.content;
-                    candidate.lang = loaded.lang;
-                    candidate.ext = loaded.ext;
-                    candidate.totalLines = loaded.totalLines;
-                    candidate.embedding = loaded.embedding;
-                    fileCache.set(candidate.path, loaded);
+            const cached = touchFileCache(candidate.path);
+            if (cached) {
+                candidate.lang ??= cached.lang;
+                candidate.ext ??= cached.ext;
+                candidate.totalLines ??= cached.totalLines;
+                candidate.embedding ??= cached.embedding;
+            }
+            const needsMetadata = candidate.lang === null || candidate.ext === null || candidate.totalLines === null;
+            const hasTextEvidence = candidate.keywordHits.size > 0 ||
+                candidate.phraseHits > 0 ||
+                candidate.fallbackTextHits > 0;
+            const needsEmbedding = candidate.embedding === null && !hasTextEvidence;
+            if (needsMetadata || needsEmbedding) {
+                const loaded = await loadFileMetadata(db, repoId, candidate.path);
+                if (!loaded) {
+                    continue;
                 }
+                candidate.lang ??= loaded.lang;
+                candidate.ext ??= loaded.ext;
+                candidate.totalLines ??= loaded.totalLines;
+                candidate.embedding ??= loaded.embedding;
+                setFileCacheEntry(candidate.path, {
+                    ...(cached ?? {}),
+                    ...loaded,
+                });
             }
             result.push(candidate);
         }
@@ -3391,11 +3526,13 @@ async function contextBundleImpl(context, params) {
         }
         for (const row of metadataRows) {
             const candidate = ensureCandidate(candidates, row.path);
-            if (row.content) {
-                candidate.content = row.content;
-                candidate.totalLines = row.content.split(/\r?\n/).length;
-                fileCache.set(row.path, {
-                    content: row.content,
+            const totalLines = coerceLineCount(row.line_count) ??
+                (row.content ? (row.content.length === 0 ? 0 : row.content.split(/\r?\n/).length) : null);
+            if (totalLines !== null) {
+                candidate.totalLines ??= totalLines;
+            }
+            if (!fileCache.has(row.path)) {
+                setFileCacheEntry(row.path, {
                     lang: row.lang,
                     ext: row.ext,
                     totalLines: candidate.totalLines,
@@ -3566,9 +3703,37 @@ async function contextBundleImpl(context, params) {
         };
     }
     const maxScore = Math.max(...prioritizedCandidates.map((candidate) => candidate.score));
+    const ensureCandidateContent = async (candidate) => {
+        if (candidate.content) {
+            return true;
+        }
+        const cached = touchFileCache(candidate.path);
+        if (cached?.content) {
+            candidate.content = cached.content;
+            candidate.lang ??= cached.lang;
+            candidate.ext ??= cached.ext;
+            candidate.totalLines ??= cached.totalLines;
+            candidate.embedding ??= cached.embedding;
+            return true;
+        }
+        const loaded = await loadFileContent(db, repoId, candidate.path);
+        if (!loaded) {
+            return false;
+        }
+        candidate.content = loaded.content;
+        candidate.lang ??= loaded.lang;
+        candidate.ext ??= loaded.ext;
+        candidate.totalLines ??= loaded.totalLines;
+        candidate.embedding ??= loaded.embedding;
+        setFileCacheEntry(candidate.path, {
+            ...(cached ?? {}),
+            ...loaded,
+        });
+        return true;
+    };
     const results = [];
     for (const candidate of prioritizedCandidates) {
-        if (!candidate.content) {
+        if (!(await ensureCandidateContent(candidate))) {
             continue;
         }
         const snippets = await db.all(`