claude-memory-admin 1.10.1 → 1.10.2

package/src/config.mjs

@@ -0,0 +1,130 @@
+// Where Claude Code keeps its files, and the path primitives that answer it the
+// same way on all three platforms.
+//
+// Everything under ~/.claude - the projects root, the agent definitions, the
+// agent memory directories, the user settings file, the output styles - moves as
+// one when CLAUDE_CONFIG_DIR is set. Hardcoding the home-relative path in each
+// module meant a machine with a custom config dir opened this tool to an empty
+// list rather than to its memory, so all of them resolve through configDir()
+// instead.
+//
+// The platform is a parameter on every function here rather than a read of
+// process.platform inside it. Only one of the three platforms is ever available
+// to develop on, and behaviour that cannot be exercised from a test is behaviour
+// nobody checks: the defaults keep the call sites unchanged and the parameter is
+// what lets test/config.test.mjs assert the Windows rules from anywhere.
+
+import os from 'node:os';
+import path from 'node:path';
+
+/** The separator rules of the named platform, rather than of the host. */
+const pathFor = (platform) => (platform === 'win32' ? path.win32 : path.posix);
+
+/**
+ * Expand a leading `~/`. Claude Code documents that form for autoMemoryDirectory
+ * and it is what people type on every platform, Windows included, so it is
+ * handled before any separator rule is applied.
+ *
+ * The join goes through pathFor rather than through path directly. A caller that
+ * names a platform is naming its separator too, and the native join would answer
+ * from the host instead: a POSIX home joined under win32 comes back rooted at no
+ * drive, which isAbsolutePath then rejects for a reason the caller never asked
+ * for.
+ */
+export function expandHome(value, { home = os.homedir(), platform = process.platform } = {}) {
+  if (typeof value !== 'string') return value;
+  if (value === '~') return home;
+  if (value.startsWith('~/') || value.startsWith('~\\')) return pathFor(platform).join(home, value.slice(2));
+  return value;
+}
+
+/**
+ * Whether a path is absolute in the sense Claude Code requires: rooted, and on
+ * Windows rooted at a drive or a UNC share.
+ *
+ * path.isAbsolute is not enough on its own. Under win32 it answers true for
+ * "/foo", which has no drive and so names nothing that can be opened. Accepting
+ * it would turn a typo into a silent read of the wrong place, which is the one
+ * failure every caller here exists to prevent.
+ */
+export function isAbsolutePath(value, platform = process.platform) {
+  if (typeof value !== 'string' || !value) return false;
+  if (platform === 'win32') {
+    return /^[A-Za-z]:[\\/]/.test(value) || /^\\\\[^\\]/.test(value) || /^\/\/[^/]/.test(value);
+  }
+  return value.startsWith('/');
+}
+
+/** A path with forward slashes, which is the only separator glob syntax knows. */
+export function toPosix(value) {
+  return typeof value === 'string' ? value.replace(/\\/g, '/') : value;
+}
+
+/**
+ * A path in the form to compare two paths by, never the form to display or store.
+ *
+ * Windows and a default macOS volume are both case-insensitive, so two spellings
+ * of one directory are one directory, and a dedup set keyed on the raw string
+ * misses the duplicate. Lowercasing is only sound for the comparison: the store
+ * ids in src/stores.mjs are derived from the path as found on disk and have to
+ * stay that way, or every id on the machine changes.
+ */
+export function canonicalPath(value, platform = process.platform) {
+  if (typeof value !== 'string' || !value) return value;
+  const resolved = toPosix(path.resolve(value));
+  return platform === 'win32' || platform === 'darwin' ? resolved.toLowerCase() : resolved;
+}
+
+export const DEFAULT_CONFIG_DIR_NAME = '.claude';
+
+/**
+ * The config directory and what decided it, in the shape src/projects.mjs
+ * already uses for the memory root: a tool that silently reads somewhere other
+ * than the default is worse than one that reads nothing, so the source travels
+ * with the path.
+ *
+ * A CLAUDE_CONFIG_DIR that is not absolute is reported and not used. Claude Code
+ * would not accept it either, so falling back to the default is the honest
+ * answer - but silently is not, hence `invalid`.
+ */
+export function configSource({ env = process.env, home = os.homedir(), platform = process.platform } = {}) {
+  const fallback = pathFor(platform).join(home, DEFAULT_CONFIG_DIR_NAME);
+  const raw = typeof env.CLAUDE_CONFIG_DIR === 'string' ? env.CLAUDE_CONFIG_DIR.trim() : '';
+  if (!raw) return { path: fallback, source: 'default', raw: null, invalid: null };
+
+  const expanded = expandHome(raw, { home, platform });
+  if (!isAbsolutePath(expanded, platform)) {
+    return {
+      path: fallback,
+      source: 'default',
+      raw,
+      invalid: `CLAUDE_CONFIG_DIR "${raw}" is not an absolute or ~/ path`,
+    };
+  }
+  return { path: expanded, source: 'env', raw, invalid: null };
+}
+
+export function configDir(options = {}) {
+  return configSource(options).path;
+}
+
+/** A path inside the config directory, such as configPath('agents'). */
+export function configPath(...segments) {
+  return path.join(configDir(), ...segments);
+}
+
+/**
+ * CLAUDE_CODE_PROJECT_DIR_NAME, which fixes the <project> directory name so that
+ * every repository launched with this config dir shares one store instead of
+ * getting a slug of its own.
+ *
+ * Claude Code documents it as being set beside CLAUDE_CONFIG_DIR, and it is only
+ * honoured here on that condition: on its own it would silently redirect a
+ * default installation to a directory Claude Code is not using.
+ */
+export function fixedProjectDirName({ env = process.env } = {}) {
+  if (typeof env.CLAUDE_CONFIG_DIR !== 'string' || !env.CLAUDE_CONFIG_DIR.trim()) return null;
+  const name = typeof env.CLAUDE_CODE_PROJECT_DIR_NAME === 'string' ? env.CLAUDE_CODE_PROJECT_DIR_NAME.trim() : '';
+  if (!name || name !== path.basename(name) || name.startsWith('.')) return null;
+  return name;
+}

package/src/cost.mjs

@@ -18,9 +18,9 @@
 //                                   prompt and are untouched by it.
 
 import fs from 'node:fs';
-import os from 'node:os';
 import path from 'node:path';
 
+import { configPath } from './config.mjs';
 import { parseFrontmatter } from './parse.mjs';
 import { writeFileAtomic } from './mutate.mjs';
 import {
@@ -31,7 +31,7 @@ import {
   settingsCandidates,
 } from './settings.mjs';
 
-export const OUTPUT_STYLES_DIR = path.join(os.homedir(), '.claude', 'output-styles');
+export const OUTPUT_STYLES_DIR = configPath('output-styles');
 
 // A full model name is documented as acceptable wherever an alias is, and this
 // tool has no business deciding which ones exist. The shape is checked so a

package/src/instructions.mjs

@@ -13,8 +13,8 @@
 // would hide exactly the problem this is meant to surface.
 
 import fs from 'node:fs';
-import os from 'node:os';
 import path from 'node:path';
+import { canonicalPath, configDir, DEFAULT_CONFIG_DIR_NAME, expandHome, toPosix } from './config.mjs';
 import { parseFrontmatter } from './parse.mjs';
 import { estimateTokens } from './stats.mjs';
 import { lookup, settingsLayers } from './settings.mjs';
@@ -133,7 +133,7 @@ export function findImports(text) {
 }
 
 function resolveImport(spec, fromFile) {
-  if (spec.startsWith('~/')) return path.join(os.homedir(), spec.slice(2));
+  if (spec.startsWith('~/') || spec.startsWith('~\\')) return expandHome(spec);
   if (path.isAbsolute(spec)) return spec;
   // Relative to the file holding the import, not to the working directory.
   return path.resolve(path.dirname(fromFile), spec);
@@ -149,7 +149,10 @@ function resolveImport(spec, fromFile) {
 export function expandImports(rootFile, rootText, { projectDir = null } = {}) {
   const files = [];
   const problems = [];
-  const seen = new Set([rootFile]);
+  // Canonical keys: on a case-insensitive filesystem two spellings of one file
+  // are one file, and a cycle spelled differently at each hop would otherwise
+  // recurse until it hit the depth limit instead of being named as a cycle.
+  const seen = new Set([canonicalPath(rootFile)]);
   let frontier = [{ file: rootFile, text: rootText, depth: 0 }];
 
   while (frontier.length) {
@@ -162,7 +165,7 @@ export function expandImports(rootFile, rootText, { projectDir = null } = {}) {
           problems.push({ kind: 'too-deep', spec, from: current.file, file: resolved });
           continue;
         }
-        if (seen.has(resolved)) {
+        if (seen.has(canonicalPath(resolved))) {
           problems.push({ kind: 'cycle', spec, from: current.file, file: resolved });
           continue;
         }
@@ -173,7 +176,7 @@ export function expandImports(rootFile, rootText, { projectDir = null } = {}) {
           continue;
         }
 
-        seen.add(resolved);
+        seen.add(canonicalPath(resolved));
         // An import resolving outside the project is the kind Claude Code asks
         // you to approve once; a declined one then stays silently disabled.
         const external = projectDir ? path.relative(projectDir, resolved).startsWith('..') : false;
@@ -342,17 +345,34 @@ function entry(file, scope, kind, { conditional = false } = {}) {
 }
 
 /** Directories from the filesystem root down to `dir`, in load order. */
+/**
+ * Every directory from the filesystem root down to `dir`, root last.
+ *
+ * Walked with dirname rather than by splitting on the separator, because the
+ * separator alone does not describe a root: a Windows path starts at `C:\` and
+ * a UNC path at `\\server\share`, and splitting either of those produces
+ * candidates that name nothing.
+ */
 function ancestors(dir) {
-  const parts = path.resolve(dir).split(path.sep);
   const out = [];
-  for (let i = 1; i < parts.length; i++) {
-    out.push(parts.slice(0, i + 1).join(path.sep) || path.sep);
+  let current = path.resolve(dir);
+  while (true) {
+    out.push(current);
+    const parent = path.dirname(current);
+    if (parent === current) break;
+    current = parent;
   }
-  return out;
+  return out.reverse();
 }
 
-function globToRegExp(pattern) {
-  // Only used for claudeMdExcludes, which matches absolute paths.
+export function globToRegExp(pattern, platform = process.platform) {
+  // Only used for claudeMdExcludes, which matches absolute paths. Glob syntax
+  // knows one separator, so a Windows path has to be compared in its forward
+  // slash form or every pattern silently matches nothing - and on the two
+  // case-insensitive platforms the comparison has to ignore case for the same
+  // reason a path does.
+  const flags = platform === 'win32' || platform === 'darwin' ? 'i' : '';
+  pattern = toPosix(pattern);
   let out = '';
   for (let i = 0; i < pattern.length; i++) {
     const char = pattern[i];
@@ -362,16 +382,26 @@ function globToRegExp(pattern) {
     } else if (char === '?') out += '[^/]';
     else out += char.replace(/[.+^${}()|[\]\\]/g, '\\$&');
   }
-  return new RegExp(`^${out}$`);
+  return new RegExp(`^${out}$`, flags);
 }
 
 /**
  * Managed policy and user scope: what loads whichever project a session starts in.
  *
- * `home` is a parameter rather than a call to os.homedir() so this is a total
+ * `home` is a parameter rather than a read of the config directory so this is a total
  * function of where it is told to look, which is what makes it testable without a
  * real home directory.
  */
+/**
+ * The user-scope directory: normally the config directory, which
+ * CLAUDE_CONFIG_DIR may have moved off the home directory entirely. A `home`
+ * passed in names its own, which is what lets the tests point the user scope at
+ * a fixture.
+ */
+function userDir(home) {
+  return home ? path.join(home, DEFAULT_CONFIG_DIR_NAME) : configDir();
+}
+
 function userCandidates(home, layers) {
   const candidates = [entry(managedClaudeMd(), 'managed', 'claude-md')];
 
@@ -382,8 +412,9 @@ function userCandidates(home, layers) {
     : null;
   if (inline) candidates.push(inline);
 
-  candidates.push(entry(path.join(home, '.claude', 'CLAUDE.md'), 'user', 'claude-md'));
-  for (const file of listRuleFiles(path.join(home, '.claude', 'rules'))) {
+  const dir = userDir(home);
+  candidates.push(entry(path.join(dir, 'CLAUDE.md'), 'user', 'claude-md'));
+  for (const file of listRuleFiles(path.join(dir, 'rules'))) {
     candidates.push(ruleEntry(file, 'user'));
   }
   return candidates;
@@ -416,14 +447,14 @@ function readExcludes(layers) {
  * one cannot drift apart in what they count or what they consider broken.
  */
 function finalize(candidates, { projectDir, excludes }) {
-  const matchers = excludes.map(globToRegExp);
+  const matchers = excludes.map((pattern) => globToRegExp(pattern));
   const excluded = [];
   const loaded = [];
   const problems = [];
 
   for (const candidate of candidates.filter(Boolean)) {
     // Managed policy cannot be excluded; that is the point of managed policy.
-    if (candidate.scope !== 'managed' && matchers.some((re) => re.test(candidate.file))) {
+    if (candidate.scope !== 'managed' && matchers.some((re) => re.test(toPosix(candidate.file)))) {
       excluded.push(candidate.file);
       continue;
     }
@@ -498,7 +529,7 @@ function finalize(candidates, { projectDir, excludes }) {
  * and walking into them would bury the one finding that matters.
  */
 export function unreferencedUserFiles(home, loadedFiles) {
-  const dir = path.join(home, '.claude');
+  const dir = userDir(home);
   // Imports arrive already absolutised by resolveImport while candidates are
   // joined from `home`, so both sides are resolved before they are compared.
   const loaded = new Set(loadedFiles.map((item) => path.resolve(item.file)));
@@ -526,7 +557,7 @@ export function unreferencedUserFiles(home, loadedFiles) {
 export function resolveInstructions(projectDir) {
   const layers = settingsLayers({ projectDir });
   const candidates = [
-    ...userCandidates(os.homedir(), layers),
+    ...userCandidates(null, layers),
     ...projectCandidates(projectDir),
   ];
   const resolved = finalize(candidates, { projectDir, excludes: readExcludes(layers) });
@@ -552,7 +583,7 @@ export function resolveInstructions(projectDir) {
  * With no project there is no boundary for an import to resolve outside of, so
  * nothing is marked external here - that check belongs to the whole-session view.
  */
-export function resolveGlobalInstructions({ home = os.homedir(), settingsFile = null } = {}) {
+export function resolveGlobalInstructions({ home = null, settingsFile = null } = {}) {
   const layers = settingsLayers(settingsFile ? { settingsFile } : {});
   const resolved = finalize(userCandidates(home, layers), { projectDir: null, excludes: readExcludes(layers) });
   resolved.problems.push(...unreferencedUserFiles(home, resolved.files));

package/src/liveSessions.mjs

@@ -4,9 +4,10 @@
 // happening right now.
 
 import fs from 'node:fs';
-import os from 'node:os';
 import path from 'node:path';
 
+import { configPath } from './config.mjs';
+
 function isAlive(pid) {
   try {
     process.kill(pid, 0);
@@ -21,7 +22,7 @@ function isAlive(pid) {
  * left behind by a process that has since exited is silently skipped rather
  * than reported as stale - the registry itself is not this app's to clean up.
  */
-export function readLiveSessions({ dir = path.join(os.homedir(), '.claude', 'sessions') } = {}) {
+export function readLiveSessions({ dir = configPath('sessions') } = {}) {
   let entries;
   try {
     entries = fs.readdirSync(dir, { withFileTypes: true });

package/src/model.mjs

@@ -15,7 +15,7 @@ import { parseIndex, parseFrontmatter, extractWikilinks } from './parse.mjs';
 import { listMemoryFiles, memoryDir, resolveProjectPath, shortLabel } from './projects.mjs';
 import { autoMemoryState } from './settings.mjs';
 import { ageInDays, estimateTokens, findDuplicates, indexStats } from './stats.mjs';
-import { memoryChecks, sessionChecks } from './checks.mjs';
+import { agentStoreChecks, memoryChecks, sessionChecks } from './checks.mjs';
 import { listSessions, originSession, retention, transcriptDir } from './sessions.mjs';
 import { rememberedPath } from './pathcache.mjs';
 
@@ -198,6 +198,7 @@ export function buildStore(store) {
     ...(health.longHooks.length
       ? [{ kind: 'long-hooks', severity: 'warn', count: health.longHooks.length, longest: health.longHooks[0] }]
       : []),
+    ...agentStoreChecks(store),
     ...memoryChecks(memories, index),
     ...sessionChecks(store, memories, sessions.retention, {
       remembered: sessions.remembered,

package/src/parse.mjs

@@ -24,6 +24,36 @@ const WIKILINK = /\[\[([^\]|]+)(?:\|([^\]]*))?\]\]/g;
 /** Leading hook separator: em dash, en dash, hyphen or colon. */
 const HOOK_SEP = /^[ \t]*(?:—|–|-|:)[ \t]*/;
 
+/**
+ * A line with its carriage return taken off, for matching only.
+ *
+ * Every classifying regex here ends in `$`, and JavaScript's `.` does not match
+ * a carriage return, so on a CRLF file each of them fails on the last character
+ * and the file parses as if it held nothing: no index entries, no frontmatter,
+ * every memory an orphan. Splitting on /\r?\n/ instead would fix the matching
+ * and break the promise that matters more - that a file is rewritten byte for
+ * byte - because it would rewrite a CRLF file as LF. So the return stays on the
+ * stored line and comes off only for the match.
+ */
+export function forMatch(line) {
+  return typeof line === 'string' && line.endsWith('\r') ? line.slice(0, -1) : line;
+}
+
+/**
+ * The carriage return a line added to this file should end with, so a new entry
+ * matches the file it joins rather than making it half one thing and half the
+ * other. Decided by what most of the file already does.
+ */
+export function fileEol(lines) {
+  let crlf = 0;
+  let lf = 0;
+  for (let i = 0; i < lines.length - 1; i++) {
+    if (lines[i].endsWith('\r')) crlf += 1;
+    else lf += 1;
+  }
+  return crlf > lf ? '\r' : '';
+}
+
 function unquote(value) {
   const trimmed = value.trim();
   if (trimmed.length >= 2) {
@@ -50,7 +80,7 @@ function readYaml(lines) {
     const withoutComment = raw.trimStart().startsWith('#') ? '' : raw;
     if (!withoutComment.trim()) continue;
 
-    const match = withoutComment.match(/^([ \t]*)([A-Za-z0-9_.\-]+):[ \t]*(.*)$/);
+    const match = forMatch(withoutComment).match(/^([ \t]*)([A-Za-z0-9_.\-]+):[ \t]*(.*)$/);
     if (!match) continue;
 
     const [, indent, key, rest] = match;
@@ -103,7 +133,7 @@ export function parseFrontmatter(text) {
   return {
     data: readYaml(block),
     raw: block.join('\n'),
-    body: lines.slice(end + 1).join('\n').replace(/^\n+/, ''),
+    body: lines.slice(end + 1).join('\n').replace(/^(?:\r?\n)+/, ''),
     hasFrontmatter: true,
   };
 }
@@ -140,12 +170,12 @@ export function parseIndex(text) {
   let section = null;
 
   lines.forEach((line, i) => {
-    const heading = line.match(/^(#{1,6})[ \t]+(.*)$/);
+    const heading = forMatch(line).match(/^(#{1,6})[ \t]+(.*)$/);
     if (heading) {
       section = heading[2].trim();
       parsedLines.push({ index: i, kind: 'heading', text: line, level: heading[1].length, section });
     } else {
-      const indexMatch = line.match(INDEX_LINE);
+      const indexMatch = forMatch(line).match(INDEX_LINE);
       // Only an unindented bullet is an index entry. An indented one is a
       // sub-bullet of surrounding prose and must not be treated as a pointer.
       if (indexMatch && indexMatch[1] === '') {
@@ -255,19 +285,23 @@ export function unwrapWikilink(text, target) {
 }
 
 function splitEntryLine(line) {
-  const match = line.match(INDEX_LINE);
+  const body = forMatch(line);
+  const match = body.match(INDEX_LINE);
   if (!match || match[1] !== '') return null;
   const file = match[3];
   const marker = `](${file})`;
-  const linkEnd = line.indexOf(marker) + marker.length;
-  const tail = line.slice(linkEnd);
+  const linkEnd = body.indexOf(marker) + marker.length;
+  const tail = body.slice(linkEnd);
   const separator = tail.match(HOOK_SEP);
   return {
-    head: line.slice(0, linkEnd),
+    head: body.slice(0, linkEnd),
     title: match[2].trim(),
     file,
     separator: separator ? separator[0] : null,
     hook: tail.replace(HOOK_SEP, ''),
+    // Carried so a rewritten line ends the way the line it replaces did. A file
+    // that comes back half CRLF and half LF is a file this tool corrupted.
+    eol: line.endsWith('\r') ? '\r' : '',
   };
 }
 
@@ -317,7 +351,7 @@ export function setIndexHook(text, index, expectedText, hook) {
 
   const separator = parts.separator || dominantSeparator(parseIndex(text));
   const before = lines[index];
-  lines[index] = trimmed ? `${parts.head}${separator}${trimmed}` : parts.head;
+  lines[index] = (trimmed ? `${parts.head}${separator}${trimmed}` : parts.head) + parts.eol;
   return { text: lines.join('\n'), before, after: lines[index] };
 }
 
@@ -396,7 +430,7 @@ export function sectionInsertIndex(parsed, section) {
 export function insertIndexEntry(text, at, line) {
   const lines = text.split('\n');
   const target = Math.max(0, Math.min(Number(at), lines.length));
-  lines.splice(target, 0, line);
+  lines.splice(target, 0, line + fileEol(lines));
   return { text: lines.join('\n'), index: target };
 }
 

package/src/pathcheck.mjs

@@ -22,7 +22,12 @@ function candidate(raw) {
   if (token.includes('://') || token.startsWith('@')) return null;
   if (GLOB_OR_SHELL.test(token)) return null;
 
+  // A memory written on Windows quotes its paths with backslashes. The project
+  // index below is keyed on forward slashes, so a token that is not normalised
+  // matches nothing and every Windows-written path is reported as missing.
+  token = token.replace(/\\/g, '/');
   token = token.replace(LINE_SUFFIX, '').replace(/[.,;:]+$/, '').replace(/\/+$/, '');
+  if (/^[A-Za-z]:\//.test(token)) return null;
   if (token.startsWith('./')) token = token.slice(2);
   if (!token || token.startsWith('/') || token.startsWith('~')) return null;
   if (token.split('/').includes('..')) return null;

package/src/projects.mjs

@@ -14,12 +14,12 @@
 // and the repository root preferred as the store's identity.
 
 import fs from 'node:fs';
-import os from 'node:os';
 import path from 'node:path';
+import { configPath, configSource, fixedProjectDirName } from './config.mjs';
 import { rememberedPath } from './pathcache.mjs';
 import { autoMemoryState, expandHome, resolveMemoryDirectory } from './settings.mjs';
 
-export const DEFAULT_ROOT = path.join(os.homedir(), '.claude', 'projects');
+export const DEFAULT_ROOT = configPath('projects');
 
 /**
  * Which store to read, and what decided it. The source travels with the path so
@@ -44,7 +44,24 @@ export function resolveRoot() {
       invalid: `autoMemoryDirectory "${configured.raw}" is ${configured.invalid}`,
     };
   }
-  return { path: DEFAULT_ROOT, source: 'default', file: null, invalid: null };
+  // CLAUDE_CONFIG_DIR moved the whole config directory, so DEFAULT_ROOT is
+  // already inside it. Naming that as the source keeps the UI from claiming a
+  // default location the user never used.
+  const config = configSource();
+  if (config.source === 'env') {
+    return { path: DEFAULT_ROOT, source: 'config-dir', file: null, invalid: config.invalid };
+  }
+  return { path: DEFAULT_ROOT, source: 'default', file: null, invalid: config.invalid };
+}
+
+/**
+ * The one project directory every repository shares when
+ * CLAUDE_CODE_PROJECT_DIR_NAME is set, or null when each repository gets a slug
+ * of its own. Callers use it to explain why one row stands for several
+ * repositories rather than to change how the row is read.
+ */
+export function fixedProjectDir() {
+  return fixedProjectDirName();
 }
 
 export function projectsRoot() {
@@ -180,12 +197,29 @@ export function storeIdentity(cwds) {
  * Candidates are verified against the filesystem so a wrong guess is reported
  * as unresolved rather than presented as fact.
  */
-function decodeSlug(slug) {
+export function slugCandidates(slug, platform = process.platform) {
+  // A Windows cwd slugifies with its drive in front ("C--Users-me-repo"), so the
+  // POSIX candidates below decode it to a path with no drive and no meaning.
+  if (platform === 'win32') {
+    const drive = /^([A-Za-z])--(.*)$/.exec(slug);
+    if (drive) {
+      const [, letter, rest] = drive;
+      return [
+        `${letter}:\\` + rest.replace(/--/g, '\\.').replace(/-/g, '\\'),
+        `${letter}:\\` + rest.replace(/-/g, '\\'),
+      ];
+    }
+  }
+
   const body = slug.replace(/^-/, '');
-  const candidates = [
+  return [
     '/' + body.replace(/--/g, '/.').replace(/-/g, '/'),
     '/' + body.replace(/-/g, '/'),
   ];
+}
+
+function decodeSlug(slug) {
+  const candidates = slugCandidates(slug);
   for (const candidate of candidates) {
     try {
       if (fs.existsSync(candidate)) return { path: candidate, verified: true };
@@ -223,7 +257,10 @@ export function resolveProjectPath(dir, slug) {
 
 /** A short label for the sidebar: the last two path segments. */
 export function shortLabel(fullPath) {
-  const parts = fullPath.split('/').filter(Boolean);
+  // Split on both separators: a store recorded on Windows is still browsed on
+  // whatever machine is reading it, and a backslash path would otherwise come
+  // out as one unsplittable segment.
+  const parts = String(fullPath).split(/[\\/]+/).filter(Boolean);
   return parts.slice(-2).join('/') || fullPath;
 }