mixdog 0.8.0 → 0.9.0

package/src/lib/mixdog-debug.cjs

@@ -99,11 +99,6 @@ function pruneStalePluginDataLogSiblings(
   return { removed, kept: candidates.length - removed };
 }
 
-/** @deprecated use pruneStalePluginDataLogSiblings */
-function prunePluginDataLogFiles(dataDir, maxFiles) {
-  return pruneStalePluginDataLogSiblings(dataDir, maxFiles);
-}
-
 const SESSION_START_CRITICAL_LOG = 'session-start-critical.log';
 const SESSION_START_CRITICAL_MAX_BYTES = 64 * 1024;
 const SESSION_START_CRITICAL_KEEP_BYTES = 64 * 1024;
@@ -154,10 +149,8 @@ module.exports = {
   isMixdogDebugEnabled,
   isStalePluginLogSibling,
   pruneStalePluginDataLogSiblings,
-  prunePluginDataLogFiles,
   appendSessionStartCriticalLog,
   sessionStartCriticalFallback,
   DEFAULT_STALE_LOG_SIBLING_MAX,
-  DEFAULT_PLUGIN_LOG_MAX_FILES: DEFAULT_STALE_LOG_SIBLING_MAX,
   SESSION_START_CRITICAL_LOG,
 };

package/src/lib/rules-builder.cjs

@@ -3,24 +3,28 @@
 /**
  * mixdog rules builder.
  *
- * Three surfaces:
- *   - buildInjectionContent              — Lead session
- *   - buildBridgeInjectionContent        — bridge session BP1 (true cross-role common)
- *   - buildBridgeRoleSpecificContent     — bridge session BP3 (role-specific instructions)
+ * Exported surfaces:
+ *   - buildSharedToolContent            — BP1: shared tool policy
+ *   - buildLeadRoleContent              — BP2: Lead role/system rules
+ *   - buildAgentRoleContent             — BP2: agent role/system rules
+ *   - buildAgentRetrievalInjectionContent — BP2: narrow read-only retrieval role
+ *   - buildLeadMetaContent              — BP3: Lead memory/meta context
+ *   - buildAgentRoleSpecificContent     — BP4-adjacent user/task data
+ *   - buildInjectionContent             — legacy joined Lead session content
  *
  * 4-BP cache layout (composeSystemPrompt):
- *   BP1 = bridge BP1 content (this file's buildBridgeInjectionContent) — every role identical
- *   BP2 = scoped role catalog (collect.mjs loadScopedRoleCatalog) — role family / self
- *   BP3 = project context + role marker + permission + role-specific instructions
- *   BP4 = task brief + memory recap (5m volatile)
+ *   BP1 = shared tool policy + compact skill manifest
+ *   BP2 = role/system rules (Lead / agent / hidden role)
+ *   BP3 = stable memory/meta context
+ *   BP4 = live user/task messages and compacted tail
  *
  * Source files (rules/):
- *   - shared/01-tool.md              — universal tool policy (Lead + bridge BP1, identical full set)
- *   - lead/00-tool-lead.md           — Lead-specific control-tower / delegation / ToolSearch guidance
+ *   - shared/01-tool.md              — universal tool policy (Lead + agent BP1, identical full set)
+ *   - lead/lead-tool.md             — Lead-specific control-tower / delegation / ToolSearch guidance
  *   - lead/01-04                     — Lead workflow / channels / team / general
  *   - output-styles/<name>.md        — Lead output style, selected by config outputStyle
- *   - bridge/00-common.md            — bridge common behavior + universal worker contract (BP1)
- *   - bridge/10..50-*.md             — per-hidden-role bodies (consumed by loadScopedRoleCatalog)
+ *   - agent/00-common.md             — agent common behavior + universal worker contract (BP2)
+ *   - agent/10..50-*.md              — per-hidden-role bodies (consumed by loadScopedRoleInstructions)
  *
  * Core memory snapshot and session recap are injected separately by
  * hooks/session-start.cjs from the memory worker (pgdata) (Lead only).
@@ -56,8 +60,96 @@ function readUnifiedConfig(dataDir) {
   return {};
 }
 
+function readAgentConfig(dataDir) {
+  const unified = readUnifiedConfig(dataDir);
+  return unified.agent && typeof unified.agent === 'object' ? unified.agent : unified;
+}
+
+const PROFILE_LANGUAGE_PROMPTS = Object.freeze({
+  en: 'English',
+  ko: 'Korean (한국어)',
+  ja: 'Japanese (日本語)',
+  'zh-Hans': 'Simplified Chinese (简体中文)',
+  'zh-Hant': 'Traditional Chinese (繁體中文)',
+  es: 'Spanish (Español)',
+  fr: 'French (Français)',
+  de: 'German (Deutsch)',
+  pt: 'Portuguese (Português)',
+  ru: 'Russian (Русский)',
+  it: 'Italian (Italiano)',
+  vi: 'Vietnamese (Tiếng Việt)',
+  th: 'Thai (ภาษาไทย)',
+  id: 'Indonesian (Bahasa Indonesia)',
+  hi: 'Hindi (हिन्दी)',
+  ar: 'Arabic (العربية)',
+  tr: 'Turkish (Türkçe)',
+  pl: 'Polish (Polski)',
+  nl: 'Dutch (Nederlands)',
+  uk: 'Ukrainian (Українська)',
+});
+
+const PROFILE_TITLE_MAX = 64;
+
+function normalizeProfileConfig(value = {}) {
+  const raw = value && typeof value === 'object' ? value : {};
+  const title = String(raw.title ?? raw.name ?? '').trim().slice(0, PROFILE_TITLE_MAX);
+  const requested = String(raw.language ?? raw.lang ?? 'system').trim();
+  const language = requested === 'system' || PROFILE_LANGUAGE_PROMPTS[requested] ? requested : 'system';
+  return { title, language };
+}
+
+function systemLocaleId(locale) {
+  let parsed = null;
+  try { parsed = new Intl.Locale(locale || ''); } catch {}
+  const language = parsed?.language || String(locale || '').split(/[-_]/)[0] || '';
+  if (language === 'zh') {
+    const script = parsed?.script;
+    const region = parsed?.region;
+    return script === 'Hant' || ['HK', 'MO', 'TW'].includes(region) ? 'zh-Hant' : 'zh-Hans';
+  }
+  return PROFILE_LANGUAGE_PROMPTS[language] ? language : null;
+}
+
+function profileLanguagePrompt(language) {
+  const selected = String(language || 'system');
+  if (selected !== 'system') {
+    return PROFILE_LANGUAGE_PROMPTS[selected]
+      ? { prompt: PROFILE_LANGUAGE_PROMPTS[selected], source: 'profile', locale: null }
+      : null;
+  }
+  const locale = Intl.DateTimeFormat().resolvedOptions().locale || '';
+  const id = systemLocaleId(locale);
+  if (!id || !PROFILE_LANGUAGE_PROMPTS[id]) return null;
+  return { prompt: PROFILE_LANGUAGE_PROMPTS[id], source: 'system-locale', locale };
+}
+
+function buildProfilePreferencesContent(dataDir) {
+  const profile = normalizeProfileConfig(readAgentConfig(dataDir).profile);
+  const lines = [];
+  if (profile.title) {
+    lines.push(`- Use "${profile.title}" when directly addressing the user (greetings, answers, questions). Do not repeat it in routine progress updates or pre-tool preambles.`);
+  }
+  const shell = process.platform === 'win32' ? 'powershell' : 'bash';
+  lines.push(`- Shell environment: ${shell}. When using shell, write commands and scripts in ${shell} syntax unless the user specifies otherwise. Keep commands, paths, symbols, and exact errors verbatim.`);
+  return lines.length ? `# Profile Preferences\n\n${lines.join('\n')}` : '';
+}
+
+function buildLanguageSection(dataDir) {
+  const profile = normalizeProfileConfig(readAgentConfig(dataDir).profile);
+  const language = profileLanguagePrompt(profile.language);
+  if (!language?.prompt) return '';
+  const source = language.source === 'system-locale' && language.locale
+    ? ` from system locale ${language.locale}`
+    : '';
+  const lines = [
+    `- Default user-facing response language${source}: ${language.prompt}. Use it for ALL user-facing prose, including pre-tool preambles, progress updates, questions, final reports, and notices. Every such message — even a single-line preamble before a tool call — MUST be written in ${language.prompt}, and in no other language. This overrides any tone implied by the output style. Switch languages only when the user writes in another language or explicitly asks you to.`,
+    `- Code identifiers, paths, commands, symbols, API names, and exact errors should remain in their original form.`,
+  ];
+  return `# Language\n\n${lines.join('\n')}`;
+}
+
 function stripFrontmatter(markdown) {
-  return String(markdown || '').replace(/^---\r?\n[\s\S]*?\r?\n---\r?\n*/, '').trim();
+  return String(markdown || '').replace(/^---[ \t]*\r?\n[\s\S]*?\r?\n---[ \t]*(?:\r?\n|$)/, '').trim();
 }
 
 function normalizeOutputStyleName(value) {
@@ -91,8 +183,8 @@ function loadOutputStyle({ PLUGIN_ROOT, DATA_DIR }) {
 }
 
 /**
- * Resolve the DATA_DIR subdir whose *.md instruction tree folds into a hidden
- * role's BP3 role-specific block, from the role's `instructionDir` metadata in
+ * Resolve the DATA_DIR subdir whose *.md instruction tree is sent as
+ * BP4-adjacent user/task data, from the role's `instructionDir` metadata in
  * defaults/hidden-roles.json. Returns null when the role declares none.
  *
  * Mirrors internal-roles.mjs getRoleInstructionDir — rules-builder is CommonJS
@@ -154,67 +246,92 @@ function composeUserAddressBullet(memoryConfig) {
   return `- User address form: ${userTitle} (use this address form exactly as written; do not combine it with any other name or honorific)`;
 }
 
-/**
- * Build the Lead injection content.
- */
-function buildInjectionContent({ PLUGIN_ROOT, DATA_DIR }) {
+function splitLeadGeneral(general, addressBullet = '') {
+  const lines = String(general || '').split(/\r?\n/);
+  const roleLines = [];
+  const metaLines = [];
+  let inMeta = false;
+  for (const line of lines) {
+    if (/language used by the user/i.test(line)) {
+      if (!inMeta) {
+        metaLines.push('# General');
+        metaLines.push('');
+        inMeta = true;
+      }
+      metaLines.push(line);
+    } else {
+      roleLines.push(line);
+    }
+  }
+  if (addressBullet) {
+    if (!inMeta) {
+      metaLines.push('# General');
+      metaLines.push('');
+      inMeta = true;
+    }
+    metaLines.push(addressBullet);
+  }
+  return {
+    role: roleLines.join('\n').trim(),
+    meta: metaLines.join('\n').trim(),
+  };
+}
+
+function buildSharedToolContent({ PLUGIN_ROOT }) {
+  const SHARED_DIR = path.join(PLUGIN_ROOT, 'rules', 'shared');
+  return readOptional(path.join(SHARED_DIR, '01-tool.md'));
+}
+
+function buildLeadRoleContent({ PLUGIN_ROOT, DATA_DIR }) {
   const RULES_DIR = path.join(PLUGIN_ROOT, 'rules');
-  const SHARED_DIR = path.join(RULES_DIR, 'shared');
   const LEAD_DIR = path.join(RULES_DIR, 'lead');
-  const HISTORY_DIR = path.join(DATA_DIR, 'history');
-
   const memoryConfig = readConfigSection(DATA_DIR, 'memory');
-  const parts = [];
-
-  // Language policy injects first — global default, applied to every Lead reply
-  // and to internal role communication regardless of locale.
-  const language = readOptional(path.join(SHARED_DIR, '00-language.md'));
-  if (language) parts.push(language);
-
+  const addressBullet = composeUserAddressBullet(memoryConfig);
   const general = readOptional(path.join(LEAD_DIR, '01-general.md'));
-  if (general) {
-    const addressBullet = composeUserAddressBullet(memoryConfig);
-    parts.push(addressBullet ? `${general}\n${addressBullet}` : general);
-  }
-
-  const tool = readOptional(path.join(SHARED_DIR, '01-tool.md'));
-  if (tool) parts.push(tool);
+  const generalSplit = splitLeadGeneral(general, addressBullet);
+  const parts = [];
 
-  const toolLead = readOptional(path.join(LEAD_DIR, '00-tool-lead.md'));
+  const toolLead = readOptional(path.join(LEAD_DIR, 'lead-tool.md'));
   if (toolLead) parts.push(toolLead);
 
+  if (generalSplit.role) parts.push(generalSplit.role);
+
   const channels = readOptional(path.join(LEAD_DIR, '02-channels.md'));
   if (channels) parts.push(channels);
 
   const workflow = readOptional(path.join(LEAD_DIR, '04-workflow.md'));
   if (workflow) parts.push(workflow);
 
+  return parts.join('\n\n');
+}
+
+function buildLeadMetaContent({ PLUGIN_ROOT, DATA_DIR }) {
+  const RULES_DIR = path.join(PLUGIN_ROOT, 'rules');
+  const LEAD_DIR = path.join(RULES_DIR, 'lead');
+  const HISTORY_DIR = path.join(DATA_DIR, 'history');
+  const memoryConfig = readConfigSection(DATA_DIR, 'memory');
+  const addressBullet = composeUserAddressBullet(memoryConfig);
+  const general = readOptional(path.join(LEAD_DIR, '01-general.md'));
+  const generalSplit = splitLeadGeneral(general, addressBullet);
+  const parts = [];
+
+  const profilePreferences = buildProfilePreferencesContent(DATA_DIR);
+  if (profilePreferences) parts.push(profilePreferences);
+
+  const languageSection = buildLanguageSection(DATA_DIR);
+  if (languageSection) parts.push(languageSection);
+
+  if (generalSplit.meta) parts.push(generalSplit.meta);
+
   const userProfile = readOptional(path.join(HISTORY_DIR, 'user.md'));
   if (userProfile) parts.push(`# User Profile\n\n${userProfile}`);
 
   const botPersona = readOptional(path.join(HISTORY_DIR, 'bot.md'));
   if (botPersona) parts.push(`# Bot Persona\n\n${botPersona}`);
 
-  // User workflow context — low-token, optional
   const userWorkflowMd = readOptional(path.join(DATA_DIR, 'user-workflow.md'));
   if (userWorkflowMd) parts.push(`# User Workflow\n\n${userWorkflowMd}`);
 
-  const userWorkflowJsonPath = path.join(DATA_DIR, 'user-workflow.json');
-  try {
-    const json = JSON.parse(fs.readFileSync(userWorkflowJsonPath, 'utf8'));
-    if (json && Array.isArray(json.roles) && json.roles.length > 0) {
-      const lines = json.roles.map(r => {
-        const name = r.name || '';
-        const preset = r.preset || '';
-        const permission = r.permission || '';
-        return `- ${name}${preset ? ` (preset: ${preset})` : ''}${permission ? ` [${permission}]` : ''}`;
-      });
-      parts.push(`# User Workflow Roles\n\n${lines.join('\n')}`);
-    }
-  } catch { /* absent or invalid — skip */ }
-
-  // Keep output style last so user/profile/workflow context cannot soften the
-  // final formatting contract for user-visible replies.
   const outputStyle = loadOutputStyle({ PLUGIN_ROOT, DATA_DIR });
   if (outputStyle) parts.push(outputStyle);
 
@@ -222,61 +339,83 @@ function buildInjectionContent({ PLUGIN_ROOT, DATA_DIR }) {
 }
 
 /**
- * BP1 — true cross-role common. Identical for every bridge role; the
- * role-specific stuff (per-event webhook instructions, per-task schedule
- * instructions, hidden role tool detail) lives in BP3 instead.
+ * Build the Lead injection content.
+ */
+function buildInjectionContent({ PLUGIN_ROOT, DATA_DIR }) {
+  const parts = [];
+
+  const tool = buildSharedToolContent({ PLUGIN_ROOT, DATA_DIR });
+  if (tool) parts.push(tool);
+
+  const role = buildLeadRoleContent({ PLUGIN_ROOT, DATA_DIR });
+  if (role) parts.push(role);
+
+  const meta = buildLeadMetaContent({ PLUGIN_ROOT, DATA_DIR });
+  if (meta) parts.push(meta);
+
+  return parts.join('\n\n');
+}
+
+/**
+ * Legacy joined agent injection. New sessions consume shared tool policy as
+ * BP1 and buildAgentRoleContent() as BP2; this export remains for older smoke
+ * tests / callers that expect the combined shape.
  *
  * @param {object} opts
  * @param {string} opts.PLUGIN_ROOT
  * @param {string} opts.DATA_DIR
  * @returns {string}
  */
-function buildBridgeInjectionContent({ PLUGIN_ROOT, DATA_DIR }) {
-  const RULES_DIR = path.join(PLUGIN_ROOT, 'rules');
-  const SHARED_DIR = path.join(RULES_DIR, 'shared');
-  const BRIDGE_DIR = path.join(RULES_DIR, 'bridge');
+function buildAgentInjectionContent({ PLUGIN_ROOT, DATA_DIR }) {
   const parts = [];
 
-  // 0. Language policy — global default, language-neutral plugin behavior.
-  const language = readOptional(path.join(SHARED_DIR, '00-language.md'));
-  if (language) parts.push(language);
-
-  // 1. Universal tool policy — same full set Lead receives.
-  const tool = readOptional(path.join(SHARED_DIR, '01-tool.md'));
+  const tool = buildSharedToolContent({ PLUGIN_ROOT, DATA_DIR });
   if (tool) parts.push(tool);
 
-  // 2. Bridge common behavior.
-  const common = readOptional(path.join(BRIDGE_DIR, '00-common.md'));
-  if (common) parts.push(common);
-
-  // 3. User-defined work-role overrides (DATA_DIR/roles/*.md). Pool-wide.
-  const rolesDir = path.join(DATA_DIR, 'roles');
-  const collected = collectMarkdownFilesRecursive(rolesDir);
-  if (collected.length > 0) {
-    collected.sort();
-    const blocks = collected.map(f => readOptional(f)).filter(Boolean);
-    if (blocks.length > 0) {
-      parts.push(['# Agent roles', '', blocks.join('\n\n')].join('\n'));
-    }
-  }
-
-  // userTitle / address form is intentionally NOT injected here — bridge
-  // workers produce tool I/O, not user-facing replies, so the persona signal
-  // only biases response language/tone without serving a purpose. Lead BP1
-  // (buildInjectionContent above) still carries it.
+  const role = buildAgentRoleContent({ PLUGIN_ROOT, DATA_DIR });
+  if (role) parts.push(role);
 
   return parts.join('\n\n');
 }
 
+function buildAgentRoleContent({ PLUGIN_ROOT, profile = 'full' }) {
+  if (String(profile || 'full') === 'retrieval') {
+    return buildAgentRetrievalInjectionContent();
+  }
+  const AGENT_DIR = path.join(PLUGIN_ROOT, 'rules', 'agent');
+  return readOptional(path.join(AGENT_DIR, '00-common.md'));
+}
+
+/**
+ * BP2 role rules for narrow hidden retrieval roles. These roles already carry a separate
+ * read-only tool schema shard, so keeping the full agent worker prefix does
+ * not improve cross-role cache reuse and only adds unrelated shell/edit/git
+ * guidance.
+ *
+ * @returns {string}
+ */
+function buildAgentRetrievalInjectionContent() {
+  return [
+    '# Tool Use',
+    '',
+    '- Batch independent read-only lookups in the same tool turn.',
+    '- Use code_graph for symbols/dependencies, grep for exact text, find/glob/list for files, and read only known paths/windows.',
+    '',
+    '# Agent Constraints',
+    '',
+    '- Read-only retrieval role: do not edit files, run shell, or use git.',
+    '- Follow the role output contract exactly; do not add handoff prose.',
+  ].join('\n');
+}
+
 /**
- * BP3 role-specific instructions. Only the calling role's own task / tool
- * detail body emits — webhook-handler gets webhooks/<all-events>/, scheduler
- * gets schedules/<all-tasks>/, hidden retrieval roles get their own tool
- * detail. Other roles return ''.
+ * BP4-adjacent role-specific data. Only the calling role's own task / tool
+ * detail body emits — webhook-handler gets webhooks/<all-events>/ and
+ * scheduler gets schedules/<all-tasks>/. Other roles return ''.
  *
  * NOTE: webhook-event narrowing (one event per call) requires the inbound
- * payload's event id at compose time; not implemented yet, so all 4 webhook
- * instructions still bake into webhook-handler BP3 for now.
+ * payload's event id at compose time; not implemented yet, so all webhook
+ * instruction files still ride with the webhook-handler user/task context.
  *
  * @param {object} opts
  * @param {string} opts.PLUGIN_ROOT
@@ -284,7 +423,7 @@ function buildBridgeInjectionContent({ PLUGIN_ROOT, DATA_DIR }) {
  * @param {string|null} opts.currentRole
  * @returns {string}
  */
-function buildBridgeRoleSpecificContent({ PLUGIN_ROOT, DATA_DIR, currentRole }) {
+function buildAgentRoleSpecificContent({ PLUGIN_ROOT, DATA_DIR, currentRole }) {
   if (!currentRole) return '';
   const parts = [];
 
@@ -298,7 +437,7 @@ function buildBridgeRoleSpecificContent({ PLUGIN_ROOT, DATA_DIR, currentRole })
     const collected = collectMarkdownFilesRecursive(dir);
     if (collected.length > 0) {
       collected.sort();
-      const blocks = collected.map(f => readOptional(f)).filter(Boolean);
+      const blocks = collected.map(f => stripFrontmatter(readOptional(f))).filter(Boolean);
       if (blocks.length > 0) {
         parts.push([`# Agent ${subdirForRole}`, '', blocks.join('\n\n')].join('\n'));
       }
@@ -309,7 +448,12 @@ function buildBridgeRoleSpecificContent({ PLUGIN_ROOT, DATA_DIR, currentRole })
 }
 
 module.exports = {
+  buildSharedToolContent,
+  buildLeadRoleContent,
+  buildLeadMetaContent,
+  buildAgentRoleContent,
   buildInjectionContent,
-  buildBridgeInjectionContent,
-  buildBridgeRoleSpecificContent,
+  buildAgentInjectionContent,
+  buildAgentRetrievalInjectionContent,
+  buildAgentRoleSpecificContent,
 };

package/src/lib/text-utils.cjs

@@ -7,7 +7,7 @@
  *
  * Single source of truth for the regex set that strips:
  *   - markdown fences, headers, list markers, bold
- *   - tool/system tags emitted by Claude Code (system-reminder, tool-use-id,
+ *   - tool/system tags from assistant transcripts (system-reminder, tool-use-id,
  *     local-command-*, etc.)
  *   - mixdog tags (channel, schedule-context, teammate-message)
  *   - URLs, emoji, "Ran X", "Process exited", and other transcript noise