@modusensus/dsh-mneme 0.6.7 → 0.6.8

package/src/mirror.js

@@ -1,170 +1,170 @@
-import { mkdirSync, readFileSync, writeFileSync, existsSync, rmSync } from "node:fs";
-import { createHash } from "node:crypto";
-import { join } from "node:path";
-
-export const TYPE_FILE = {
-  preference: "preferences.md",
-  project: "projects.md",
-  decision: "decisions.md",
-  history: "history.md",
-  summary: "summary.md"
-};
-
-const ESCAPE = /([\\`*_[\]{}()#+.!|>~-])/g;
-const UNESCAPE = new RegExp("\\\\" + ESCAPE.source, "g");
-
-function esc(text) {
-  return String(text).replace(ESCAPE, "\\$1");
-}
-
-function unescape(text) {
-  return String(text).replace(UNESCAPE, "$1");
-}
-
-function renderMemory(m) {
-  // last-rendered digest baseline: sha256(title \x00 content). service.js
-  // compares the file hash against this to tell "untouched by a human" (machine
-  // write wins) apart from a real human edit, so a not-yet-re-rendered store
-  // update is not misread as a concurrent human edit.
-  const digest = createHash("sha256")
-    .update(`${m.title}\x00${m.content}`)
-    .digest("hex");
-  const lines = [];
-  lines.push(`## ${esc(m.title)}`);
-  // v0.6.2 tag line: entity_attrs-backed tags (attached by the service as
-  // `entityTags`) rendered as `#tag` space-separated under the title. No tags
-  // → no line. Legacy `- **标签**:` metadata below keeps the memories.tags
-  // column (still written by save/update) readable.
-  const entityTags = Array.isArray(m.entityTags) ? m.entityTags : [];
-  if (entityTags.length) {
-    lines.push("");
-    lines.push(entityTags.map((t) => `#${esc(t)}`).join(" "));
-  }
-  lines.push("");
-  lines.push(`- **ID**: \`${m.id}\``);
-  lines.push(`- **类型**: ${m.type}`);
-  lines.push(`- **重要性**: ${m.importance}`);
-  lines.push(`- **标签**: ${m.tags.map((t) => `\`${esc(t)}\``).join(" ")}`);
-  lines.push(`- **更新时间**: ${m.updated_at}`);
-  if (m.source) lines.push(`- **来源**: ${esc(m.source)}`);
-  lines.push("");
-  lines.push(`<!-- mirror-digest: ${digest} -->`);
-  lines.push(m.content);
-  lines.push("");
-  lines.push("---");
-  lines.push("");
-  return lines.join("\n");
-}
-
-export function createMirror(dir) {
-  mkdirSync(dir, { recursive: true });
-
-  function filePath(type) {
-    const name = TYPE_FILE[type];
-    return name ? join(dir, name) : undefined;
-  }
-
-  /**
-   * Parse a mirror file back into {id, title, content} entries for human edits.
-   * Entries are anchored on "- **ID**: `...`" lines that are followed by the
-   * "- **类型**:" metadata line (structural entry head): each entry's block
-   * spans from its ID line up to the next ID line (or end of file). The block
-   * head (the ID line plus the generated metadata run) and the trailing
-   * structural "---" separator are stripped; everything in between is the entry
-   * body, so user content containing "---", metadata-like lines, or even a
-   * machine-format "- **ID**: `x`" line is preserved. The title is the "## "
-   * heading preceding the ID line.
-   */
-  function readHumanEdits(type = undefined) {
-    const types = type ? [type] : Object.keys(TYPE_FILE);
-    const edits = [];
-    for (const t of types) {
-      const file = filePath(t);
-      if (!file || !existsSync(file)) continue;
-      const text = readFileSync(file, "utf8").replace(/\r\n/g, "\n");
-      // Anchor on the ID line only when it is a structural entry head: the
-      // machine-rendered ID line is always followed by the "- **类型**:" line.
-      // A body line like "- **ID**: `x`" is not, so it never splits the block
-      // or produces a ghost entry.
-      const anchors = [...text.matchAll(/^- \*\*ID\*\*: `([^`]+)`\n- \*\*类型\*\*:/gm)];
-      let prevEnd = 0;
-      for (let i = 0; i < anchors.length; i++) {
-        const anchor = anchors[i];
-        const blockStart = anchor.index;
-        const blockEnd = i + 1 < anchors.length ? anchors[i + 1].index : text.length;
-
-        // Title: last "## " heading before this ID line (file header region /
-        // previous block tail). Body headings of earlier entries come before
-        // the structural "---" + "## " of this entry, so the last match wins.
-        const titleMatches = [...text.slice(prevEnd, blockStart).matchAll(/^## (.+)$/gm)];
-        const titleMatch = titleMatches[titleMatches.length - 1];
-
-        // Body: the ID line and the generated metadata run are structural head;
-        // everything after them up to the trailing "---" separator is the body.
-        let body = text
-          .slice(blockStart, blockEnd)
-          .replace(/^- \*\*ID\*\*: `[^`]+`\n?/, "")
-          .replace(/^(- \*\*(类型|重要性|标签|更新时间|来源)\*\*:.*\n?)+/, "")
-          .replace(/^<!-- mirror-digest: [a-f0-9]+ -->\n?/m, "");
-        const separators = [...body.matchAll(/^---\s*$/gm)];
-        const lastSep = separators[separators.length - 1];
-        if (lastSep) body = body.slice(0, lastSep.index);
-        body = body.trim();
-
-        // The machine-written "更新时间" line records the store's updated_at at
-        // render time — the version token for detecting a concurrent store write
-        // during a three-way merge of human edits (see service.syncMirror).
-        const block = text.slice(blockStart, blockEnd);
-        const updatedMatch = block.match(/- \*\*更新时间\*\*: ([^\n]+)/);
-        const digestMatch = block.match(/<!-- mirror-digest: ([a-f0-9]+) -->/);
-        edits.push({
-          id: anchor[1],
-          title: titleMatch ? unescape(titleMatch[1]).trim() : undefined,
-          content: body,
-          updated_at: updatedMatch ? updatedMatch[1].trim() : undefined,
-          digest: digestMatch ? digestMatch[1] : undefined
-        });
-
-        const lineEnd = text.indexOf("\n", blockStart);
-        prevEnd = lineEnd === -1 ? text.length : lineEnd + 1;
-      }
-    }
-    return edits;
-  }
-
-  function sync(memories) {
-    const byType = {};
-    for (const m of memories) {
-      (byType[m.type] ??= []).push(m);
-    }
-    // Per-type physical outcomes (audit peer D): a failed write for one type
-    // must not abort the whole render. Each type is written (or pruned) in its
-    // own try/catch and the result reported so the caller can persist per-type
-    // committed/failed receipts — a file that was already written is a real
-    // physical commit even when a sibling type errors.
-    const results = {};
-    for (const type of Object.keys(TYPE_FILE)) {
-      try {
-        const file = filePath(type);
-        const items = (byType[type] ?? [])
-          .slice()
-          .sort((a, b) => (a.updated_at < b.updated_at ? 1 : -1));
-        if (items.length === 0) {
-          // no memories of this type: drop any stale mirror file so deleted
-          // memories do not "resurrect" via readHumanEdits
-          rmSync(file, { force: true });
-        } else {
-          const header = `# ${TYPE_FILE[type]} — dsh-mneme 镜像\n\n<!-- 手工编辑此文件会被合并回记忆库（人工优先）。 -->\n\n`;
-          const body = items.map(renderMemory).join("\n");
-          writeFileSync(file, header + body, "utf8");
-        }
-        results[type] = { ok: true };
-      } catch (error) {
-        results[type] = { ok: false, error: error?.message ?? String(error) };
-      }
-    }
-    return results;
-  }
-
-  return { filePath, sync, readHumanEdits };
-}
+import { mkdirSync, readFileSync, writeFileSync, existsSync, rmSync } from "node:fs";
+import { createHash } from "node:crypto";
+import { join } from "node:path";
+
+export const TYPE_FILE = {
+  preference: "preferences.md",
+  project: "projects.md",
+  decision: "decisions.md",
+  history: "history.md",
+  summary: "summary.md"
+};
+
+const ESCAPE = /([\\`*_[\]{}()#+.!|>~-])/g;
+const UNESCAPE = new RegExp("\\\\" + ESCAPE.source, "g");
+
+function esc(text) {
+  return String(text).replace(ESCAPE, "\\$1");
+}
+
+function unescape(text) {
+  return String(text).replace(UNESCAPE, "$1");
+}
+
+function renderMemory(m) {
+  // last-rendered digest baseline: sha256(title \x00 content). service.js
+  // compares the file hash against this to tell "untouched by a human" (machine
+  // write wins) apart from a real human edit, so a not-yet-re-rendered store
+  // update is not misread as a concurrent human edit.
+  const digest = createHash("sha256")
+    .update(`${m.title}\x00${m.content}`)
+    .digest("hex");
+  const lines = [];
+  lines.push(`## ${esc(m.title)}`);
+  // v0.6.2 tag line: entity_attrs-backed tags (attached by the service as
+  // `entityTags`) rendered as `#tag` space-separated under the title. No tags
+  // → no line. Legacy `- **标签**:` metadata below keeps the memories.tags
+  // column (still written by save/update) readable.
+  const entityTags = Array.isArray(m.entityTags) ? m.entityTags : [];
+  if (entityTags.length) {
+    lines.push("");
+    lines.push(entityTags.map((t) => `#${esc(t)}`).join(" "));
+  }
+  lines.push("");
+  lines.push(`- **ID**: \`${m.id}\``);
+  lines.push(`- **类型**: ${m.type}`);
+  lines.push(`- **重要性**: ${m.importance}`);
+  lines.push(`- **标签**: ${m.tags.map((t) => `\`${esc(t)}\``).join(" ")}`);
+  lines.push(`- **更新时间**: ${m.updated_at}`);
+  if (m.source) lines.push(`- **来源**: ${esc(m.source)}`);
+  lines.push("");
+  lines.push(`<!-- mirror-digest: ${digest} -->`);
+  lines.push(m.content);
+  lines.push("");
+  lines.push("---");
+  lines.push("");
+  return lines.join("\n");
+}
+
+export function createMirror(dir) {
+  mkdirSync(dir, { recursive: true });
+
+  function filePath(type) {
+    const name = TYPE_FILE[type];
+    return name ? join(dir, name) : undefined;
+  }
+
+  /**
+   * Parse a mirror file back into {id, title, content} entries for human edits.
+   * Entries are anchored on "- **ID**: `...`" lines that are followed by the
+   * "- **类型**:" metadata line (structural entry head): each entry's block
+   * spans from its ID line up to the next ID line (or end of file). The block
+   * head (the ID line plus the generated metadata run) and the trailing
+   * structural "---" separator are stripped; everything in between is the entry
+   * body, so user content containing "---", metadata-like lines, or even a
+   * machine-format "- **ID**: `x`" line is preserved. The title is the "## "
+   * heading preceding the ID line.
+   */
+  function readHumanEdits(type = undefined) {
+    const types = type ? [type] : Object.keys(TYPE_FILE);
+    const edits = [];
+    for (const t of types) {
+      const file = filePath(t);
+      if (!file || !existsSync(file)) continue;
+      const text = readFileSync(file, "utf8").replace(/\r\n/g, "\n");
+      // Anchor on the ID line only when it is a structural entry head: the
+      // machine-rendered ID line is always followed by the "- **类型**:" line.
+      // A body line like "- **ID**: `x`" is not, so it never splits the block
+      // or produces a ghost entry.
+      const anchors = [...text.matchAll(/^- \*\*ID\*\*: `([^`]+)`\n- \*\*类型\*\*:/gm)];
+      let prevEnd = 0;
+      for (let i = 0; i < anchors.length; i++) {
+        const anchor = anchors[i];
+        const blockStart = anchor.index;
+        const blockEnd = i + 1 < anchors.length ? anchors[i + 1].index : text.length;
+
+        // Title: last "## " heading before this ID line (file header region /
+        // previous block tail). Body headings of earlier entries come before
+        // the structural "---" + "## " of this entry, so the last match wins.
+        const titleMatches = [...text.slice(prevEnd, blockStart).matchAll(/^## (.+)$/gm)];
+        const titleMatch = titleMatches[titleMatches.length - 1];
+
+        // Body: the ID line and the generated metadata run are structural head;
+        // everything after them up to the trailing "---" separator is the body.
+        let body = text
+          .slice(blockStart, blockEnd)
+          .replace(/^- \*\*ID\*\*: `[^`]+`\n?/, "")
+          .replace(/^(- \*\*(类型|重要性|标签|更新时间|来源)\*\*:.*\n?)+/, "")
+          .replace(/^<!-- mirror-digest: [a-f0-9]+ -->\n?/m, "");
+        const separators = [...body.matchAll(/^---\s*$/gm)];
+        const lastSep = separators[separators.length - 1];
+        if (lastSep) body = body.slice(0, lastSep.index);
+        body = body.trim();
+
+        // The machine-written "更新时间" line records the store's updated_at at
+        // render time — the version token for detecting a concurrent store write
+        // during a three-way merge of human edits (see service.syncMirror).
+        const block = text.slice(blockStart, blockEnd);
+        const updatedMatch = block.match(/- \*\*更新时间\*\*: ([^\n]+)/);
+        const digestMatch = block.match(/<!-- mirror-digest: ([a-f0-9]+) -->/);
+        edits.push({
+          id: anchor[1],
+          title: titleMatch ? unescape(titleMatch[1]).trim() : undefined,
+          content: body,
+          updated_at: updatedMatch ? updatedMatch[1].trim() : undefined,
+          digest: digestMatch ? digestMatch[1] : undefined
+        });
+
+        const lineEnd = text.indexOf("\n", blockStart);
+        prevEnd = lineEnd === -1 ? text.length : lineEnd + 1;
+      }
+    }
+    return edits;
+  }
+
+  function sync(memories) {
+    const byType = {};
+    for (const m of memories) {
+      (byType[m.type] ??= []).push(m);
+    }
+    // Per-type physical outcomes (audit peer D): a failed write for one type
+    // must not abort the whole render. Each type is written (or pruned) in its
+    // own try/catch and the result reported so the caller can persist per-type
+    // committed/failed receipts — a file that was already written is a real
+    // physical commit even when a sibling type errors.
+    const results = {};
+    for (const type of Object.keys(TYPE_FILE)) {
+      try {
+        const file = filePath(type);
+        const items = (byType[type] ?? [])
+          .slice()
+          .sort((a, b) => (a.updated_at < b.updated_at ? 1 : -1));
+        if (items.length === 0) {
+          // no memories of this type: drop any stale mirror file so deleted
+          // memories do not "resurrect" via readHumanEdits
+          rmSync(file, { force: true });
+        } else {
+          const header = `# ${TYPE_FILE[type]} — dsh-mneme 镜像\n\n<!-- 手工编辑此文件会被合并回记忆库（人工优先）。 -->\n\n`;
+          const body = items.map(renderMemory).join("\n");
+          writeFileSync(file, header + body, "utf8");
+        }
+        results[type] = { ok: true };
+      } catch (error) {
+        results[type] = { ok: false, error: error?.message ?? String(error) };
+      }
+    }
+    return results;
+  }
+
+  return { filePath, sync, readHumanEdits };
+}

package/src/parser/tag.js

@@ -1,59 +1,59 @@
-/**
- * Tag parser (v0.6.2). Given a memory's content, extract explicit hashtags of
- * the form `#标签`. The allowed character set is `[a-zA-Z0-9_一-龥-]+` (ASCII
- * letters/digits/underscore, CJK characters, and hyphen), so `#linux`,
- * `#deepseek` and `#考研` are all valid tags.
- *
- * Rules:
- *   - multiple tags on one line are all extracted, in source order;
- *   - duplicates collapse to the first occurrence;
- *   - tags longer than MAX_TAG_LENGTH (20) characters are dropped as illegal;
- *   - edge hyphens (leading/trailing, e.g. a sentence-terminating `#tag-`) are
- *     stripped; a tag that collapses to empty is dropped;
- *   - markdown `## heading` never matches (the second `#`/space is not in the
- *     allowed set).
- *
- * Pure module: no store, no side effects — persistence (store.setMemoryTags)
- * is a separate step. sanitizeTags shares the same validation so LLM-extracted
- * tag arrays (autoDream tag-extractor) are filtered identically.
- */
-
-export const MAX_TAG_LENGTH = 20;
-
-// `一` (U+4E00) .. `龥` (U+9FA5) is the CJK Unified Ideographs block (the
-// rule's `一-龥`). Matching requires a tag character right after `#` so a bare
-// `#` and markdown `## heading` are naturally excluded (#/space not in set).
-const TAG_RE = /#[a-zA-Z0-9_一-龥-]+/g;
-const EDGE_HYPHEN = /^-+|-+$/g;
-const VALID_TAG = /^[a-zA-Z0-9_一-龥-]+$/;
-
-/** Validate + dedupe an array of tag strings. Non-strings, blanks, over-long
- *  and illegal-character tags are dropped (fail-safe). A leading `#` is
- *  tolerated so raw LLM output can be passed through directly. */
-export function sanitizeTags(tags) {
-  if (!Array.isArray(tags)) return [];
-  const seen = new Set();
-  const out = [];
-  for (const t of tags) {
-    if (typeof t !== "string") continue;
-    let tag = t.trim();
-    if (tag.startsWith("#")) tag = tag.slice(1);
-    tag = tag.replace(EDGE_HYPHEN, "");
-    if (!tag || tag.length > MAX_TAG_LENGTH) continue;
-    if (!VALID_TAG.test(tag)) continue;
-    if (seen.has(tag)) continue; // 去重（保留首次出现顺序）
-    seen.add(tag);
-    out.push(tag);
-  }
-  return out;
-}
-
-/** @returns {string[]} validated tags in source order, deduplicated. */
-export function parseTags(content) {
-  if (typeof content !== "string" || content.length === 0) return [];
-  const raw = [];
-  TAG_RE.lastIndex = 0;
-  let m;
-  while ((m = TAG_RE.exec(content)) !== null) raw.push(m[0]);
-  return sanitizeTags(raw);
-}
+/**
+ * Tag parser (v0.6.2). Given a memory's content, extract explicit hashtags of
+ * the form `#标签`. The allowed character set is `[a-zA-Z0-9_一-龥-]+` (ASCII
+ * letters/digits/underscore, CJK characters, and hyphen), so `#linux`,
+ * `#deepseek` and `#考研` are all valid tags.
+ *
+ * Rules:
+ *   - multiple tags on one line are all extracted, in source order;
+ *   - duplicates collapse to the first occurrence;
+ *   - tags longer than MAX_TAG_LENGTH (20) characters are dropped as illegal;
+ *   - edge hyphens (leading/trailing, e.g. a sentence-terminating `#tag-`) are
+ *     stripped; a tag that collapses to empty is dropped;
+ *   - markdown `## heading` never matches (the second `#`/space is not in the
+ *     allowed set).
+ *
+ * Pure module: no store, no side effects — persistence (store.setMemoryTags)
+ * is a separate step. sanitizeTags shares the same validation so LLM-extracted
+ * tag arrays (autoDream tag-extractor) are filtered identically.
+ */
+
+export const MAX_TAG_LENGTH = 20;
+
+// `一` (U+4E00) .. `龥` (U+9FA5) is the CJK Unified Ideographs block (the
+// rule's `一-龥`). Matching requires a tag character right after `#` so a bare
+// `#` and markdown `## heading` are naturally excluded (#/space not in set).
+const TAG_RE = /#[a-zA-Z0-9_一-龥-]+/g;
+const EDGE_HYPHEN = /^-+|-+$/g;
+const VALID_TAG = /^[a-zA-Z0-9_一-龥-]+$/;
+
+/** Validate + dedupe an array of tag strings. Non-strings, blanks, over-long
+ *  and illegal-character tags are dropped (fail-safe). A leading `#` is
+ *  tolerated so raw LLM output can be passed through directly. */
+export function sanitizeTags(tags) {
+  if (!Array.isArray(tags)) return [];
+  const seen = new Set();
+  const out = [];
+  for (const t of tags) {
+    if (typeof t !== "string") continue;
+    let tag = t.trim();
+    if (tag.startsWith("#")) tag = tag.slice(1);
+    tag = tag.replace(EDGE_HYPHEN, "");
+    if (!tag || tag.length > MAX_TAG_LENGTH) continue;
+    if (!VALID_TAG.test(tag)) continue;
+    if (seen.has(tag)) continue; // 去重（保留首次出现顺序）
+    seen.add(tag);
+    out.push(tag);
+  }
+  return out;
+}
+
+/** @returns {string[]} validated tags in source order, deduplicated. */
+export function parseTags(content) {
+  if (typeof content !== "string" || content.length === 0) return [];
+  const raw = [];
+  TAG_RE.lastIndex = 0;
+  let m;
+  while ((m = TAG_RE.exec(content)) !== null) raw.push(m[0]);
+  return sanitizeTags(raw);
+}

package/src/parser/wiki-link.js

@@ -1,38 +1,38 @@
-/**
- * Wiki-Link parser (v0.6.1). Given a memory's content, extract explicit
- * cross-memory links of the form [[target]] or [[显示|target]]:
- *   [[target]]      → { display: "target", target: "target" }
- *   [[显示|target]] → { display: "显示",  target: "target" }
- *
- * Unclosed / empty-target / multi-pipe / bracket-nested markers are treated as
- * illegal and ignored. Pure module: no store, no side effects — resolution is a
- * separate step (resolveWikiLink) that needs a store handle.
- */
-
-/** @returns {{display: string, target: string}[]} in source order. */
-export function parseWikiLinks(content) {
-  if (typeof content !== "string" || content.length === 0) return [];
-  const links = [];
-  // [^\[\]]* keeps a single match from crossing `]]`; an unclosed `[[` never
-  // matches, and `[[a [[b]] c]]` only yields the inner `[[b]]`.
-  const re = /\[\[([^\[\]]*)\]\]/g;
-  let m;
-  while ((m = re.exec(content)) !== null) {
-    const inner = m[1];
-    const parts = inner.split("|");
-    if (parts.length > 2) continue; // 多管道 → 非法，忽略
-    const rawTarget = (parts.length === 2 ? parts[1] : parts[0]).trim();
-    if (!rawTarget) continue; // 空目标 → 非法，忽略
-    const rawDisplay = parts[0].trim();
-    links.push({ display: rawDisplay || rawTarget, target: rawTarget });
-  }
-  return links;
-}
-
-/** Resolve a wiki-link target title to a memory row via a case-insensitive
- *  exact title match (store.findByTitle). Returns undefined when absent or the
- *  store exposes no such lookup. */
-export function resolveWikiLink(store, title) {
-  if (!store || typeof title !== "string" || !title.trim()) return undefined;
-  return store.findByTitle?.(title.trim());
-}
+/**
+ * Wiki-Link parser (v0.6.1). Given a memory's content, extract explicit
+ * cross-memory links of the form [[target]] or [[显示|target]]:
+ *   [[target]]      → { display: "target", target: "target" }
+ *   [[显示|target]] → { display: "显示",  target: "target" }
+ *
+ * Unclosed / empty-target / multi-pipe / bracket-nested markers are treated as
+ * illegal and ignored. Pure module: no store, no side effects — resolution is a
+ * separate step (resolveWikiLink) that needs a store handle.
+ */
+
+/** @returns {{display: string, target: string}[]} in source order. */
+export function parseWikiLinks(content) {
+  if (typeof content !== "string" || content.length === 0) return [];
+  const links = [];
+  // [^\[\]]* keeps a single match from crossing `]]`; an unclosed `[[` never
+  // matches, and `[[a [[b]] c]]` only yields the inner `[[b]]`.
+  const re = /\[\[([^\[\]]*)\]\]/g;
+  let m;
+  while ((m = re.exec(content)) !== null) {
+    const inner = m[1];
+    const parts = inner.split("|");
+    if (parts.length > 2) continue; // 多管道 → 非法，忽略
+    const rawTarget = (parts.length === 2 ? parts[1] : parts[0]).trim();
+    if (!rawTarget) continue; // 空目标 → 非法，忽略
+    const rawDisplay = parts[0].trim();
+    links.push({ display: rawDisplay || rawTarget, target: rawTarget });
+  }
+  return links;
+}
+
+/** Resolve a wiki-link target title to a memory row via a case-insensitive
+ *  exact title match (store.findByTitle). Returns undefined when absent or the
+ *  store exposes no such lookup. */
+export function resolveWikiLink(store, title) {
+  if (!store || typeof title !== "string" || !title.trim()) return undefined;
+  return store.findByTitle?.(title.trim());
+}