mixdog 0.9.18 → 0.9.19

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.18",
+  "version": "0.9.19",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -26,7 +26,8 @@
   "files": [
     "README.md",
     "scripts/",
-    "!scripts/bench/cache-hit-*.json",
+    "!scripts/bench/",
+    "!scripts/recall-bench-*.txt",
     "src/",
     "vendor/"
   ],

package/scripts/output-style-smoke.mjs

@@ -97,7 +97,7 @@ try {
 
   mkdirSync(join(dataDir, 'output-styles'), { recursive: true });
   writeFileSync(join(dataDir, 'mixdog-config.json'), JSON.stringify({
-    agent: { profile: { title: '재영님', language: 'system' } },
+    agent: { profile: { title: '홍길동님', language: 'system' } },
     outputStyle: 'custom-smoke',
   }));
   writeFileSync(join(dataDir, 'output-styles', 'custom-smoke.md'), '---\nname: custom-smoke\n---\n\n# Custom Output Style\n\ncustom smoke style\n');
@@ -105,7 +105,7 @@ try {
   assert(customRules.includes('# Custom Output Style'), 'configured outputStyle must select custom style');
   assert(!customRules.includes('Mixdog default — the most detailed of the three styles'), 'custom outputStyle should not append default style');
   const profileMeta = rulesBuilder.buildLeadMetaContent({ PLUGIN_ROOT: join(root, 'src'), DATA_DIR: dataDir });
-  assert(profileMeta.includes('Use "재영님" when directly addressing the user'), 'profile title must inject into Lead BP3 meta');
+  assert(profileMeta.includes('Use "홍길동님" when directly addressing the user'), 'profile title must inject into Lead BP3 meta');
   assert(profileMeta.includes('do not repeat it in routine progress updates or pre-tool preambles'), 'profile title must not encourage title in preambles');
   assert(/Default user-facing response language from system locale/.test(profileMeta), 'system profile language must resolve from system locale');
   assert(profileMeta.includes('pre-tool preambles (even single-line)'), 'profile language must cover pre-tool preambles');

package/scripts/recall-bench-cases.json

@@ -0,0 +1,11 @@
+[
+  { "id": "browse-last", "label": "query-less recent browse (period=last)", "args": { "period": "last", "limit": 10 }, "expect": "browse" },
+  { "id": "period-24h", "label": "period window 24h", "args": { "period": "24h", "limit": 10 }, "expect": "browse" },
+  { "id": "period-7d", "label": "period window 7d", "args": { "period": "7d", "limit": 10 }, "expect": "browse" },
+  { "id": "scope-all", "label": "all-scope query", "args": { "query": "recall", "projectScope": "all", "limit": 10 }, "expect": "results" },
+  { "id": "recent-dryrun-delta", "label": "recent-work topical: dryrun delta sum", "args": { "query": "드라이런 델타 합산", "limit": 10 }, "expect": { "kind": "results", "topNContains": ["드라이런"], "topN": 5 } },
+  { "id": "recent-remote-singleton", "label": "recent-work topical: remote singleton seat", "args": { "query": "remote 싱글톤 좌석", "limit": 10 }, "expect": { "kind": "results", "topNContains": ["싱글톤"], "topN": 5 } },
+  { "id": "recency-today", "label": "query+period recency: today's work (24h)", "args": { "query": "오늘 진행한 작업", "period": "24h", "limit": 10 }, "expect": { "kind": "results", "recencyOrdered": true } },
+  { "id": "negative-projectaa", "label": "negative: every row must literally mention the term (no filler)", "args": { "query": "ProjectAA 밸런스", "limit": 10 }, "expect": { "kind": "allContain", "allContain": ["projectaa"] } },
+  { "id": "xlang-embedding-crash", "label": "cross-language: embedding worker crash (strict topN)", "args": { "query": "embedding worker crash", "limit": 10 }, "expect": { "kind": "results", "topNContains": ["embedding"], "topN": 3 } }
+]

package/scripts/recall-bench.mjs

@@ -113,13 +113,87 @@ function scoreTopNContains(lines, substrings, n) {
   return { perSubstring, hitAtN, mrr, n };
 }
 
-function evaluateCase(kase, { count, ms, isError }, quality) {
+// Parse leading "[YYYY-MM-DD HH:MM]" timestamps from result lines and check
+// they are non-increasing (newest-first). Returns null when fewer than 2
+// lines carry a parseable timestamp (nothing to order).
+function scoreRecencyOrdered(lines) {
+  const tsRe = /^\s*\[(\d{4}-\d{2}-\d{2}[ T]\d{2}:\d{2})/;
+  const headerRe = /^\s*##\s/;
+  const parse = (raw) => Date.parse(raw.replace(' ', 'T'));
+  // Non-increasing check within a single ordered list of {raw,ts}. Returns
+  // the first {prev,cur} pair where cur is newer than prev, else null.
+  const firstBreak = (stamps) => {
+    for (let i = 1; i < stamps.length; i++) {
+      if (Number.isFinite(stamps[i].ts) && Number.isFinite(stamps[i - 1].ts) && stamps[i].ts > stamps[i - 1].ts) {
+        return { prev: stamps[i - 1].raw, cur: stamps[i].raw };
+      }
+    }
+    return null;
+  };
+  const hasSessions = lines.some((l) => headerRe.test(l));
+  if (!hasSessions) {
+    // Ungrouped: single global non-increasing check over all timestamped lines.
+    const stamps = [];
+    for (const line of lines) {
+      const m = tsRe.exec(line);
+      if (m) stamps.push({ raw: m[1], ts: parse(m[1]) });
+    }
+    if (stamps.length < 2) return { parsed: stamps.length, groups: 0, ordered: true, firstViolation: null };
+    const b = firstBreak(stamps);
+    return { parsed: stamps.length, groups: 0, ordered: b === null, firstViolation: b };
+  }
+  // Session-grouped: partition timestamped lines by their "## session" header.
+  // Timestamps only compared WITHIN a group; groups themselves must descend by
+  // their first line's timestamp.
+  const groups = [];
+  let cur = null;
+  for (const line of lines) {
+    if (headerRe.test(line)) { cur = { stamps: [] }; groups.push(cur); continue; }
+    const m = tsRe.exec(line);
+    if (m && cur) cur.stamps.push({ raw: m[1], ts: parse(m[1]) });
+  }
+  const nonEmpty = groups.filter((g) => g.stamps.length);
+  const parsed = nonEmpty.reduce((s, g) => s + g.stamps.length, 0);
+  let firstViolation = null;
+  for (const g of nonEmpty) {
+    const b = firstBreak(g.stamps);
+    if (b) { firstViolation = { ...b, scope: 'within-session' }; break; }
+  }
+  if (!firstViolation) {
+    const heads = nonEmpty.map((g) => g.stamps[0]);
+    const b = firstBreak(heads);
+    if (b) firstViolation = { ...b, scope: 'across-sessions' };
+  }
+  return { parsed, groups: nonEmpty.length, ordered: firstViolation === null, firstViolation };
+}
+
+// Score expect.allContain: every result line must contain at least one of the
+// given substrings (case-insensitive). Returns offending lines (matching none).
+// Use for negative cases where rows legitimately mention the term.
+function scoreAllContain(lines, substrings) {
+  const needles = substrings.map((s) => String(s || '').toLowerCase()).filter(Boolean);
+  const offenders = lines.filter((line) => {
+    const l = line.toLowerCase();
+    return !needles.some((n) => l.includes(n));
+  });
+  return { needles: substrings, offenders, ok: offenders.length === 0 };
+}
+
+function evaluateCase(kase, { count, ms, isError }, quality, recency, allContain) {
   const warnings = [];
   if (isError) warnings.push('error result');
   if (ms > WARN_LATENCY_MS) warnings.push(`latency ${ms}ms > ${WARN_LATENCY_MS}ms`);
   if ((kase.expect === 'browse' || kase.expect === 'idlookup') && count === 0) {
     warnings.push('0 results for a browse/id-lookup case (expected data present)');
   }
+  if (kase.expect === 'empty' && count > 0) {
+    warnings.push(`expected empty but got ${count} result(s) — possible filler/unrelated match`);
+  }
+  if (allContain) {
+    for (const line of allContain.offenders) {
+      warnings.push(`allContain miss: result line contains none of [${allContain.needles.join(', ')}]: "${line}"`);
+    }
+  }
   if (quality) {
     for (const p of quality.perSubstring) {
       if (!p.hit) {
@@ -129,6 +203,10 @@ function evaluateCase(kase, { count, ms, isError }, quality) {
       }
     }
   }
+  if (recency && !recency.ordered && recency.firstViolation) {
+    const v = recency.firstViolation;
+    warnings.push(`recencyOrdered violation (${v.scope || 'global'}): ${v.cur} newer than prior ${v.prev}`);
+  }
   const status = warnings.length ? 'WARN' : 'PASS';
   return { status, warnings };
 }
@@ -156,7 +234,10 @@ async function runCase(memoryModule, kase) {
   const topNContains = expectObj && Array.isArray(expectObj.topNContains) ? expectObj.topNContains : null;
   const cutoffN = expectObj && Number.isInteger(expectObj.topN) ? expectObj.topN : 5;
   const quality = topNContains ? scoreTopNContains(resultLines(text), topNContains, cutoffN) : null;
-  const evalResult = evaluateCase({ ...kase, expect: expectKind }, { count, ms, isError }, quality);
+  const recency = expectObj && expectObj.recencyOrdered ? scoreRecencyOrdered(resultLines(text)) : null;
+  const allContainNeedles = expectObj && Array.isArray(expectObj.allContain) ? expectObj.allContain : null;
+  const allContain = allContainNeedles ? scoreAllContain(resultLines(text), allContainNeedles) : null;
+  const evalResult = evaluateCase({ ...kase, expect: expectKind }, { count, ms, isError }, quality, recency, allContain);
   return {
     id: kase.id,
     label: kase.label,
@@ -167,6 +248,8 @@ async function runCase(memoryModule, kase) {
     errMsg,
     top3: topN(text, 3),
     quality,
+    recency,
+    allContain,
     status: evalResult.status,
     warnings: evalResult.warnings,
   };
@@ -182,6 +265,9 @@ function printCase(row) {
       process.stdout.write(`    substr "${p.needle}" -> rank ${p.rank ?? 'none'}${p.hit ? '' : '  (miss)'}\n`);
     }
   }
+  if (row.recency) {
+    process.stdout.write(`  recency: ${row.recency.ordered ? 'ordered' : 'OUT-OF-ORDER'} (${row.recency.parsed} timestamped lines)\n`);
+  }
   if (row.top3.length) {
     for (const line of row.top3) process.stdout.write(`    - ${line}\n`);
   } else {
@@ -218,6 +304,7 @@ function printSummary(rows) {
 async function main() {
   const casesPath = argValue('cases', DEFAULT_CASES_PATH);
   const jsonMode = hasFlag('json');
+  const strict = hasFlag('strict');
   const cases = loadCases(casesPath);
 
   let memoryModule;
@@ -262,6 +349,8 @@ async function main() {
 
   const hardErrors = rows.filter((r) => r.isError);
   if (hardErrors.length) process.exitCode = 1;
+  // --strict: any WARN row fails the run. Default behavior (errors-only) unchanged.
+  if (strict && rows.some((r) => r.status === 'WARN')) process.exitCode = 1;
 }
 
 main().catch((e) => {

package/scripts/tool-efficiency-diag.mjs

@@ -66,7 +66,10 @@ for (const target of ['heavy-worker', 'worker']) {
       }
       if (n === 'grep' && ['content', 'content_with_context'].includes(String(a.output_mode || ''))) {
         grepContent++;
-        if (a['-C'] != null || a['-A'] != null || a['-B'] != null) grepContentWithCtx++;
+        // Implicit context: output_mode content_with_context carries context
+        // without explicit -C/-A/-B flags.
+        if (a['-C'] != null || a['-A'] != null || a['-B'] != null
+          || String(a.output_mode) === 'content_with_context') grepContentWithCtx++;
         const gpath = typeof a.path === 'string' ? a.path : JSON.stringify(a.path || '');
         for (let j = i + 1; j < Math.min(i + 4, tools.length); j++) {
           const u = tools[j];

package/scripts/tool-smoke.mjs

@@ -508,8 +508,8 @@ if (!/^read 2\b/m.test(String(readRegionBatchOut))
   || (String(readRegionBatchOut).match(/scripts\/smoke\.mjs \[full\] \[ok\]/g) || []).length < 2
   || !/1→import \{ spawnSync \}/.test(String(readRegionBatchOut))
   || !/3→import \{ fileURLToPath \}/.test(String(readRegionBatchOut))
-  || !/pass offset:2 to continue/.test(String(readRegionBatchOut))
-  || !/pass offset:4 to continue/.test(String(readRegionBatchOut))) {
+  || !/(pass offset:2 to continue|ONE window: offset:2, limit:\d+)/.test(String(readRegionBatchOut))
+  || !/(pass offset:4 to continue|ONE window: offset:4, limit:\d+)/.test(String(readRegionBatchOut))) {
   throw new Error(`read region batch must preserve both requested spans:\n${readRegionBatchOut}`);
 }
 
@@ -534,6 +534,46 @@ if (readStringifiedLineErr || readStringifiedLineArgs.path[0].offset !== 7 || re
   throw new Error(`read guard must losslessly convert legacy line/context inside stringified arrays to offset/limit: err=${readStringifiedLineErr} args=${JSON.stringify(readStringifiedLineArgs)}`);
 }
 
+// Absorb shape 1: region array + top-level offset/limit → top-level becomes
+// the default window for regions that lack their own; no hard error.
+const readRegionPlusTopLevelArgs = {
+  path: [{ path: 'scripts/smoke.mjs', offset: 3, limit: 4 }, { path: 'scripts/smoke.mjs' }],
+  offset: 0,
+  limit: 2,
+};
+const readRegionPlusTopLevelErr = validateBuiltinArgs('read', readRegionPlusTopLevelArgs);
+if (readRegionPlusTopLevelErr
+  || 'offset' in readRegionPlusTopLevelArgs || 'limit' in readRegionPlusTopLevelArgs
+  || readRegionPlusTopLevelArgs.path[0].offset !== 3 || readRegionPlusTopLevelArgs.path[0].limit !== 4
+  || readRegionPlusTopLevelArgs.path[1].offset !== 0 || readRegionPlusTopLevelArgs.path[1].limit !== 2) {
+  throw new Error(`read guard must absorb region-array + top-level offset/limit: err=${readRegionPlusTopLevelErr} args=${JSON.stringify(readRegionPlusTopLevelArgs)}`);
+}
+
+// Absorb shape 2: parallel offset/limit as JSON-stringified arrays with path[]
+// → zipped into per-file region objects (pairwise recovery), no int error.
+const readZipWindowArgs = {
+  path: ['scripts/smoke.mjs', 'scripts/smoke.mjs'],
+  offset: '[0, 5]',
+  limit: '[2, 3]',
+};
+const readZipWindowErr = validateBuiltinArgs('read', readZipWindowArgs);
+if (readZipWindowErr || !Array.isArray(readZipWindowArgs.path)
+  || readZipWindowArgs.path[0].offset !== 0 || readZipWindowArgs.path[0].limit !== 2
+  || readZipWindowArgs.path[1].offset !== 5 || readZipWindowArgs.path[1].limit !== 3
+  || 'offset' in readZipWindowArgs || 'limit' in readZipWindowArgs) {
+  throw new Error(`read guard must zip stringified offset/limit arrays onto path[]: err=${readZipWindowErr} args=${JSON.stringify(readZipWindowArgs)}`);
+}
+
+// Absorb shape 3: code_graph file/files as a JSON-stringified array → parsed to
+// a real array before lookup (dispatched into files[]).
+const cgStringifiedFileArgs = { mode: 'symbols', file: JSON.stringify(['a.mjs', 'b.mjs']) };
+const cgStringifiedFileErr = validateBuiltinArgs('code_graph', cgStringifiedFileArgs);
+if (cgStringifiedFileErr || 'file' in cgStringifiedFileArgs
+  || !Array.isArray(cgStringifiedFileArgs.files)
+  || cgStringifiedFileArgs.files[0] !== 'a.mjs' || cgStringifiedFileArgs.files[1] !== 'b.mjs') {
+  throw new Error(`code_graph guard must parse JSON-stringified file array: err=${cgStringifiedFileErr} args=${JSON.stringify(cgStringifiedFileArgs)}`);
+}
+
 const graphOut = await executeCodeGraphTool('code_graph', {
   mode: 'symbols',
   file: 'scripts/smoke.mjs',
@@ -549,6 +589,17 @@ if (!/# symbol_search executeBuiltinTool\b/.test(String(graphSymbolBatchOut)) ||
   throw new Error(`code_graph symbol_search symbols[] batch execution failed:\n${graphSymbolBatchOut}`);
 }
 
+// Absorb shape 3 (real dispatch): file as a JSON-stringified array batches per
+// file instead of hitting "file not found: [...]".
+const graphStringifiedFileOut = await executeCodeGraphTool('code_graph', {
+  mode: 'symbols',
+  file: JSON.stringify(['scripts/smoke.mjs']),
+}, root);
+if (/file not found/.test(String(graphStringifiedFileOut))
+  || !/binding|spawnSync|symbol/i.test(String(graphStringifiedFileOut))) {
+  throw new Error(`code_graph must parse JSON-stringified file array before lookup:\n${graphStringifiedFileOut}`);
+}
+
 const graphMissingFileOut = await executeCodeGraphTool('code_graph', {
   mode: 'symbols',
   file: 'src/runtime/loop.mjs',
@@ -634,6 +685,33 @@ if (!/^Error[\s:[]/.test(String(shellTimeoutOut)) || !/\[timeout: 500ms\b/.test(
   throw new Error(`bash timeout must be milliseconds and classified as Error:\n${shellTimeoutOut}`);
 }
 
+// Auto-promotion: a sync foreground command still running past the (soft)
+// promotion budget is detached into a tracked background task and returns the
+// same task_id envelope as explicit async — the caller never pre-chose async.
+// Shrink the budget via MIXDOG_SHELL_AUTO_BACKGROUND_MS so the smoke stays fast.
+const _priorAutoBgBudget = process.env.MIXDOG_SHELL_AUTO_BACKGROUND_MS;
+process.env.MIXDOG_SHELL_AUTO_BACKGROUND_MS = '800';
+let shellAutoPromoteOut;
+try {
+  shellAutoPromoteOut = await executeBuiltinTool('shell', {
+    command: 'Start-Sleep -Seconds 6; Write-Output tool-smoke-autopromote-done',
+    cwd: root,
+    timeout: 30_000,
+    shell: 'powershell',
+  }, root);
+} finally {
+  if (_priorAutoBgBudget === undefined) delete process.env.MIXDOG_SHELL_AUTO_BACKGROUND_MS;
+  else process.env.MIXDOG_SHELL_AUTO_BACKGROUND_MS = _priorAutoBgBudget;
+}
+if (!/auto-backgrounded/i.test(String(shellAutoPromoteOut)) || !/task_id:\s*\S+/i.test(String(shellAutoPromoteOut))) {
+  throw new Error(`shell auto-promotion must return a background task envelope (task_id + auto-backgrounded):\n${shellAutoPromoteOut}`);
+}
+// Clean up the promoted job so it doesn't outlive the smoke as an orphan.
+const _autoPromoteTaskId = (/task_id:\s*(\S+)/i.exec(String(shellAutoPromoteOut)) || [])[1];
+if (_autoPromoteTaskId) {
+  try { await executeBuiltinTool('task', { action: 'cancel', task_id: _autoPromoteTaskId }, root); } catch {}
+}
+
 const legacyEscapedAlternationErr = validateBuiltinArgs('grep', { pattern: 'state\\.items\\.map\\|items\\.map', path: root });
 if (legacyEscapedAlternationErr) {
   throw new Error(`grep legacy \\| alternation should be accepted: ${legacyEscapedAlternationErr}`);
@@ -1217,8 +1295,10 @@ setInternalToolsProvider({
     if (workerToolNames.includes('tool_search')) {
       throw new Error(`agent session schema must not expose deferred tool_search: ${workerToolNames.join(', ')}`);
     }
-    if (workerToolNames.includes('shell')) {
-      throw new Error(`read-write agent session schema must not expose shell: ${workerToolNames.join(', ')}`);
+    for (const name of ['shell', 'task']) {
+      if (!workerToolNames.includes(name)) {
+        throw new Error(`read-write agent session schema must expose ${name} for self-verification: ${workerToolNames.join(', ')}`);
+      }
     }
     for (const name of ['skills_list', 'skill_view', 'skill_execute']) {
       if (workerToolNames.includes(name)) {
@@ -1266,7 +1346,7 @@ setInternalToolsProvider({
     const fullTools = (fullAgentSession.tools || []).map((tool) => tool?.name).filter(Boolean);
     const publicExploreTools = (publicExploreSession.tools || []).map((tool) => tool?.name).filter(Boolean);
     const expectedReadTools = ['code_graph', 'find', 'glob', 'list', 'grep', 'read', 'explore', 'search', 'web_fetch', 'Skill'];
-    const expectedWriteTools = ['code_graph', 'find', 'glob', 'list', 'grep', 'read', 'apply_patch', 'explore', 'search', 'web_fetch', 'Skill'];
+    const expectedWriteTools = ['code_graph', 'find', 'glob', 'list', 'grep', 'read', 'apply_patch', 'shell', 'task', 'explore', 'search', 'web_fetch', 'Skill'];
     if (JSON.stringify(readTools) !== JSON.stringify(expectedReadTools)) {
       throw new Error(`read agent schema must be fixed allow-list: expected=${expectedReadTools.join(', ')} actual=${readTools.join(', ')}`);
     }
@@ -1276,20 +1356,17 @@ setInternalToolsProvider({
     if (readTools.includes('tool_search') || writeTools.includes('tool_search')) {
       throw new Error(`agent session fixed schemas must omit tool_search: read=${readTools.join(', ')} write=${writeTools.join(', ')}`);
     }
-    if (readTools.includes('shell') || writeTools.includes('shell')) {
-      throw new Error(`read/read-write agent schemas must omit shell: read=${readTools.join(', ')} write=${writeTools.join(', ')}`);
+    if (readTools.includes('shell')) {
+      throw new Error(`read agent schema must omit shell: read=${readTools.join(', ')}`);
     }
     for (const name of ['shell', 'apply_patch', 'task']) {
       if (readTools.includes(name)) {
         throw new Error(`read agent schema must omit non-read tool ${name}: read=${readTools.join(', ')}`);
       }
     }
-    for (const name of ['apply_patch', 'task']) {
-      if (name === 'apply_patch' && !writeTools.includes(name)) {
-        throw new Error(`read-write agent schema must preserve apply_patch: write=${writeTools.join(', ')}`);
-      }
-      if (name !== 'apply_patch' && writeTools.includes(name)) {
-        throw new Error(`read-write agent schema must omit non-edit tool ${name}: write=${writeTools.join(', ')}`);
+    for (const name of ['apply_patch', 'shell', 'task']) {
+      if (!writeTools.includes(name)) {
+        throw new Error(`read-write agent schema must preserve ${name}: write=${writeTools.join(', ')}`);
       }
     }
     for (const name of ['memory', 'recall', 'reply']) {
@@ -1334,7 +1411,7 @@ setInternalToolsProvider({
   try {
     const resumed = await resumeSession(resumeAgentSession.id, 'full');
     const resumedTools = (resumed?.tools || []).map((tool) => tool?.name).filter(Boolean);
-    const expectedWriteTools = ['code_graph', 'find', 'glob', 'list', 'grep', 'read', 'apply_patch', 'explore', 'search', 'web_fetch', 'Skill'];
+    const expectedWriteTools = ['code_graph', 'find', 'glob', 'list', 'grep', 'read', 'apply_patch', 'shell', 'task', 'explore', 'search', 'web_fetch', 'Skill'];
     if (JSON.stringify(resumedTools) !== JSON.stringify(expectedWriteTools)) {
       throw new Error(`resumed read-write agent schema must keep fixed allow-list: expected=${expectedWriteTools.join(', ')} actual=${resumedTools.join(', ')}`);
     }
@@ -1522,7 +1599,7 @@ const readPathDescription = readPathSchema.description || '';
 if (!/File path/i.test(readPathDescription) || !/\{path,offset,limit\}\[\]/i.test(readPathDescription) || !/Pass arrays directly/i.test(readPathDescription) || !/legacy recovery only/i.test(readPathDescription)) {
   throw new Error('read schema must keep directory-vs-file guidance');
 }
-if (!/Dirs use list/i.test((BUILTIN_TOOLS.find((tool) => tool.name === 'read')?.description) || '')) {
+if (!/Not for directory listing/i.test((BUILTIN_TOOLS.find((tool) => tool.name === 'read')?.description) || '')) {
   throw new Error('read description must keep directory-vs-file guidance');
 }
 const readTool = BUILTIN_TOOLS.find((tool) => tool.name === 'read');
@@ -1533,7 +1610,7 @@ const readArrayItemAnyOf = readArraySchema?.items?.anyOf || [];
 if (!readArrayItemAnyOf.some((entry) => entry?.type === 'object' && entry?.properties?.offset && entry?.properties?.limit)) {
   throw new Error('read schema must expose array-of-region objects for batched spans');
 }
-if (/line\+context/i.test(readDescription) || !/numeric offset\+limit/i.test(readDescription) || !/Batch paths\/regions as real arrays/i.test(readDescription) || !/grep content_with_context or code_graph anchors/i.test(readDescription)) {
+if (/line\+context/i.test(readDescription) || !/numeric offset\+limit/i.test(readDescription) || !/Batch paths\/regions as real arrays/i.test(readDescription)) {
   throw new Error('read description must expose offset/limit as the single window form');
 }
 if (readProps.line || readProps.context) {
@@ -1596,9 +1673,6 @@ if (codeGraphSymbolSearchErr) {
 if (!/code structure\/flow/i.test(codeGraphDescription) || !/symbols\/references\/calls\/deps/i.test(codeGraphDescription)) {
   throw new Error('code_graph description must stay structure-oriented and name its symbol modes');
 }
-if (!/before text grep/i.test(codeGraphDescription)) {
-  throw new Error('code_graph description must steer symbol/caller lookups to structure lookup before grep');
-}
 if (!/repo-local/i.test(codeGraphDescription) || !/NOT web search|not web/i.test(codeGraphDescription)) {
   throw new Error('code_graph description must mark itself repo-local (not web search)');
 }
@@ -1853,23 +1927,23 @@ const grepPathDescription = grepTool?.inputSchema?.properties?.path?.description
 const grepGlobDescription = grepTool?.inputSchema?.properties?.glob?.description || '';
 const grepOutputModeDescription = grepTool?.inputSchema?.properties?.output_mode?.description || '';
 const grepHeadLimitDescription = grepTool?.inputSchema?.properties?.head_limit?.description || '';
-if (!/synonyms in pattern\[\]/i.test(grepPatternDescription) || !/OR in ONE grep/i.test(grepPatternDescription) || !/serial rewording/i.test(grepPatternDescription) || !/equivalent repeats/i.test(grepPatternDescription) || !/Narrowest file\/dir/i.test(grepPathDescription) || !/broad scopes return paths first/i.test(grepPathDescription) || !/refine from returned paths/i.test(grepPathDescription)) {
+if (!/Array = variants in one call/i.test(grepPatternDescription) || !/Known file or dir/i.test(grepPathDescription)) {
   throw new Error('grep schema must keep compact pattern/path guidance');
 }
-if (!/same grep/i.test(grepGlobDescription) || !/no follow-up grep/i.test(grepGlobDescription) || !/equivalent scope changes/i.test(grepGlobDescription)) {
-  throw new Error('grep glob schema must steer narrowing into the same grep call');
+if (!/narrow scope/i.test(grepGlobDescription)) {
+  throw new Error('grep glob schema must describe scope narrowing');
 }
-if (!/Broad scope: files_with_matches\/count/i.test(grepOutputModeDescription) || !/Narrow scope: content_with_context/i.test(grepOutputModeDescription) || !/answer from/i.test(grepOutputModeDescription) || !/skip read/i.test(grepOutputModeDescription) || !/span is not shown/i.test(grepOutputModeDescription)) {
-  throw new Error('grep output_mode schema must steer content_with_context away from follow-up reads');
+if (!/files_with_matches\/count/i.test(grepOutputModeDescription) || !/content_with_context/i.test(grepOutputModeDescription)) {
+  throw new Error('grep output_mode schema must name its output shapes');
 }
-if (grepTool?.inputSchema?.properties?.head_limit?.minimum !== 0 || !/keep small/i.test(grepHeadLimitDescription)) {
+if (grepTool?.inputSchema?.properties?.head_limit?.minimum !== 0 || !/Max results/i.test(grepHeadLimitDescription)) {
   throw new Error('grep head_limit schema must keep locator caps explicit');
 }
 if (grepTool?.inputSchema?.properties?.type) {
   throw new Error('grep type schema must stay hidden; prefer glob for extension narrowing');
 }
-if (!/code flow before text search/i.test(codeGraphProps.mode?.description || '') || !/one anchor is enough/i.test(codeGraphProps.symbols?.description || '') || !/one call/i.test(codeGraphProps.symbols?.description || '')) {
-  throw new Error('code_graph schema fields must steer away from repeated lookup loops');
+if (!/Repo-local/i.test(codeGraphProps.mode?.description || '') || !/one call/i.test(codeGraphProps.symbols?.description || '')) {
+  throw new Error('code_graph schema fields must stay compact and repo-local');
 }
 
 const longToolSearchText = compactToolSearchDescription(`${patchDescription}\n${patchDescription}`);

package/src/agents/debugger/AGENT.md

@@ -8,7 +8,7 @@ Root-cause analysis agent.
 
 Smallest confirmed cause chain before fixes. Return likely cause, evidence
 (`file:line`), smallest next check/fix. Mark confirmed facts vs inferences;
-avoid broad speculation. Fragments only; no narration, no report bloat.
+avoid broad speculation.
 
-Converge, don't sweep: when new reads stop adding evidence, report the best
-cause chain so far — a partial/blocked report is a valid completion.
+Converge, don't sweep: when new evidence stops accruing, report the best
+cause chain so far.

package/src/agents/heavy-worker/AGENT.md

@@ -5,16 +5,13 @@ permission: read-write
 # Heavy Worker
 Broad implementation agent.
 
-Bounded slices; smallest coherent change, not rewrite. Stop: unclear scope,
-growing blast radius, or Lead-only verification.
+Bounded slices; smallest coherent change, not rewrite. Stop when scope is
+unclear or the blast radius grows.
 
-EDIT-FIRST DISCIPLINE. Survey the slice with ONE batched read/grep round, then
-patch the first bounded piece — edit incrementally, don't read exhaustively.
-NEVER "one more confirming read": a plausible anchor means the next call is
-`apply_patch`. Repeated read-only turns without an edit = stalling; on a
-runtime reminder, patch the piece you understand or report blocked — a valid
-completion. Self-check comes AFTER edits; deep verification is Lead's.
+EDIT-FIRST DISCIPLINE. Survey the slice with one batched retrieval round,
+then patch the first bounded piece — edit incrementally, don't read
+exhaustively. Repeated read-only turns without an edit = stalling; on a
+runtime reminder, patch the piece you understand or report blocked.
 
-Minimal checks + how-to-verify. Hand off outcome as fragments anchored to
-`file:line`; no narration or bloat.
+Self-verify edits with shell (targeted test/build).
 

package/src/agents/maintainer/AGENT.md

@@ -7,5 +7,4 @@ permission: read-write
 Maintenance and cleanup agent.
 
 Smallest coherent change; no scope growth. Repeated read-only turns without
-an edit = stalling; patch or return blocked — a blocked report is a valid
-completion. Hand off as fragments (`file:line`); no narration.
+an edit = stalling; patch or return blocked.

package/src/agents/reviewer/AGENT.md

@@ -8,5 +8,4 @@ Regression/risk review agent.
 
 Find actionable correctness/regression/security/verification risks. Findings
 first, severity-ordered, one line with `file:line`; skip non-risky nits. If
-clean, one line + only material residual risk. Fragments only; no narration,
-no report bloat.
+clean, one line + only material residual risk.

package/src/agents/worker/AGENT.md

@@ -7,13 +7,10 @@ Scoped implementation agent.
 
 Smallest scoped change; no drive-by cleanup. Stop when done/blocked.
 
-EDIT-FIRST DISCIPLINE. Brief anchors (`file:line`) are pre-verified — trust and
-patch. No anchor: locate with AT MOST 1-2 reads, then edit — NEVER "one more
-confirming read"; if you know the file and the change, the next call is
-`apply_patch`. Repeated read-only turns without an edit = stalling; on a
-runtime reminder, patch now or return blocked with what's missing — a blocked
-report is a valid completion. Threshold is "plausible", not "proven";
-self-check comes AFTER the edit, and Lead/Reviewer own final verification.
+EDIT-FIRST DISCIPLINE. Brief anchors (`file:line`) are pre-verified — trust
+and patch. No anchor: one locating batch, then `apply_patch`. Repeated
+read-only turns without an edit = stalling; on a runtime reminder, patch now
+or return blocked. Threshold is "plausible", not "proven".
 
-Hand off outcome as fragments anchored to `file:line`; no narration or bloat.
+Self-verify edits with shell (node --check/test).
 

package/src/defaults/agents.json

@@ -7,6 +7,7 @@
       "description": "Filesystem navigation agent invoked by the `explore` MCP tool",
       "invokedBy": "explore",
       "toolSchemaProfile": "read",
+      "schemaAllowedTools": ["grep", "find", "glob", "code_graph"],
       "kind": "retrieval",
       "permission": "read",
       "stallCap": { "idleSeconds": 240, "toolRunningSeconds": 180 }
@@ -54,6 +55,7 @@
       "description": "Scheduled-task executor invoked by scheduler tick",
       "invokedBy": "scheduler",
       "toolSchemaProfile": "read-write-search",
+      "schemaAllowedTools": ["code_graph", "find", "glob", "list", "grep", "read", "apply_patch", "search", "web_fetch"],
       "kind": "maintenance",
       "permission": "read-write",
       "inboundEvent": true,
@@ -67,6 +69,7 @@
       "description": "Webhook payload handler invoked by inbound webhook events",
       "invokedBy": "webhook",
       "toolSchemaProfile": "read-write-search",
+      "schemaAllowedTools": ["code_graph", "find", "glob", "list", "grep", "read", "apply_patch", "search", "web_fetch"],
       "kind": "maintenance",
       "permission": "read-write",
       "inboundEvent": true,

package/src/mixdog-session-runtime.mjs

@@ -315,7 +315,7 @@ async function profiledImport(label, spec, { optional = false } = {}) {
     throw error;
   }
 }
-const outputStyleStatus = (dataDir = STANDALONE_DATA_DIR) => outputStyleStatusRaw(STANDALONE_ROOT, dataDir || STANDALONE_DATA_DIR);
+const outputStyleStatus = (dataDir = STANDALONE_DATA_DIR, opts = {}) => outputStyleStatusRaw(STANDALONE_ROOT, dataDir || STANDALONE_DATA_DIR, opts);
 // Workflow/agent pack loaders bound to this runtime's root/data layout.
 const {
   listWorkflowPacks,
@@ -325,6 +325,7 @@ const {
   activeWorkflowSummary,
   loadAgentDefinition,
   workflowContextBlock,
+  activeWorkflowContext,
 } = createWorkflowHelpers({
   rootDir: STANDALONE_ROOT,
   dataDir: STANDALONE_DATA_DIR,
@@ -918,6 +919,16 @@ export async function createMixdogSessionRuntime({
           stopRemote('superseded-by-newer-remote-session');
           emitRemoteStateChange(false, 'superseded');
         }
+        // Symmetric acquire: the worker took the bridge seat (boot make-before-
+        // break or activate). Flip remote ON via the same side-state a user
+        // /remote toggles — remoteEnabled + transcript writer — but WITHOUT
+        // re-invoking channel start (the worker is already running; startRemote
+        // would re-fork/activate). Idempotent: no-op when already enabled.
+        if (msg?.params?.state === 'acquired' && !remoteEnabled) {
+          remoteEnabled = true;
+          ensureRemoteTranscriptWriter();
+          emitRemoteStateChange(true, 'acquired');
+        }
         return;
       }
       if (msg?.method !== 'notifications/claude/channel') return;
@@ -1339,18 +1350,24 @@ export async function createMixdogSessionRuntime({
       await resolveMissingRouteModelForFirstTurn();
       requireModelRoute();
       bootProfile('session:create:route-ready', { ms: (performance.now() - startedAt).toFixed(1) });
-      await refreshRouteEffort();
+      // refreshRouteEffort (effort/model-meta) and loadCoreMemoryContext (memory
+      // files) are independent — refreshRouteEffort only touches route effort/
+      // display fields, never provider/model that the memory load reads — so run
+      // them concurrently instead of serially on the boot path.
+      const [, coreMemoryContext] = await Promise.all([
+        refreshRouteEffort(),
+        loadCoreMemoryContext(),
+      ]);
       bootProfile('session:create:effort-ready', { ms: (performance.now() - startedAt).toFixed(1) });
       const providerImpl = reg.getProvider(route.provider);
       if (!providerImpl) {
         throw new Error(`Provider "${route.provider}" is not configured.`);
       }
       bootProfile('session:create:provider-ready', { ms: (performance.now() - startedAt).toFixed(1) });
-      const coreMemoryContext = await loadCoreMemoryContext();
       if (closeRequested) throw new Error('runtime is closing');
       const dataDir = cfgMod.getPluginData?.() || STANDALONE_DATA_DIR;
-      const workflow = activeWorkflowSummary(config, dataDir);
-      const workflowContext = workflowContextBlock(config, dataDir);
+      // Load the active WORKFLOW.md pack once for both summary + context block.
+      const { summary: workflow, context: workflowContext } = activeWorkflowContext(config, dataDir);
       const sessionOpts = {
         provider: route.provider,
         model: route.model,
@@ -2991,7 +3008,7 @@ export async function createMixdogSessionRuntime({
       let selectedRoute = resolveRoute(config, requested);
       await reg.initProviders(ensureProviderEnabled(config, selectedRoute.provider));
       const modelMeta = await lookupModelMeta(selectedRoute.provider, selectedRoute.model);
-      // Guard (A안 / strict reject): a free-text string must never become the
+      // Guard (option A / strict reject): a free-text string must never become the
       // lead model. When the caller did not pin a provider, the route did not
       // match a preset, and the model is unknown to both the live provider
       // catalog (modelMeta is still the {id,provider} placeholder) and the

package/src/rules/agent/00-core.md

@@ -1,9 +1,8 @@
 # Agent Constraints
 
 - Use English for agent task communication.
-- ONE TURN = ONE BATCH: emit all independent tool calls of a step in the same
-  turn; a lone read-only call that could have ridden the previous turn is a
-  wasted round-trip.
+- ONE TURN = ONE BATCH: a read-only call that could have ridden the prior
+  turn is a wasted round-trip.
 - NEVER PREAMBLE: no status/progress narration, "I will..." setup, or any text
   before tool calls — call needed tools immediately; emit text only for the
   final handoff after tool work is done.
@@ -15,6 +14,8 @@
   searched); state only what changed and where to look. Verification =
   command + result in one line. Same fact twice = delete one.
 - Handoff cap ~30 lines unless `Deliver:` raises it — a ceiling, not a target.
+- A blocked or partial report is a valid completion: state done/missing/
+  blocker in fragments — never keep retrieving to avoid reporting.
 - Banned as pure cost: report headings, markdown tables (unless requested),
   prose narration, raw logs/tool traces, speculative next-checks, restated
   brief, articles/politeness.