mixdog 0.9.4 → 0.9.6

package/scripts/recall-bench.mjs

@@ -68,27 +68,67 @@ function textOfResult(result) {
 }
 
 function countEntryLines(text) {
-  const t = String(text || '').trim();
-  if (!t || t === '(no results)' || t === '(no valid ids)') return 0;
-  return t.split('\n').filter((line) => line.trim() && !line.startsWith('[recall truncated') && !line.startsWith('note:')).length;
+  return resultLines(text).length;
 }
 
 function topN(text, n = 3, maxLen = 140) {
-  const t = String(text || '').trim();
-  if (!t || t === '(no results)') return [];
-  return t.split('\n')
-    .filter((line) => line.trim() && !line.startsWith('[recall truncated') && !line.startsWith('note:'))
+  return resultLines(text)
     .slice(0, n)
     .map((line) => (line.length > maxLen ? `${line.slice(0, maxLen - 1)}…` : line));
 }
 
-function evaluateCase(kase, { count, ms, isError }) {
+// Full ordered list of result lines (no slice), used for topNContains rank
+// scoring. Same filtering rule as countEntryLines/topN.
+function resultLines(text) {
+  const t = String(text || '').trim();
+  if (!t || t === '(no results)' || t === '(no valid ids)') return [];
+  return t.split('\n').filter((line) => {
+    const trimmed = line.trim();
+    return trimmed
+      && !trimmed.startsWith('[recall truncated')
+      && !trimmed.startsWith('note:')
+      && !/^\[[^\]]+\]$/.test(trimmed);
+  });
+}
+
+// Score expect.topNContains: for each expected substring, find the 1-indexed
+// rank of the first result line containing it (case-insensitive), over the
+// FULL result list (not just the topN cutoff) so we can tell "missed
+// entirely" from "present but ranked below cutoff". hit@N/MRR are then
+// computed against the topN cutoff.
+function scoreTopNContains(lines, substrings, n) {
+  const lower = lines.map((l) => l.toLowerCase());
+  const perSubstring = substrings.map((needle) => {
+    const hay = String(needle || '').toLowerCase();
+    let rank = null;
+    for (let i = 0; i < lower.length; i++) {
+      if (hay && lower[i].includes(hay)) { rank = i + 1; break; }
+    }
+    const hit = rank !== null && rank <= n;
+    return { needle, rank, hit, rr: hit ? 1 / rank : 0 };
+  });
+  const total = perSubstring.length || 1;
+  const hitAtN = perSubstring.reduce((s, p) => s + (p.hit ? 1 : 0), 0) / total;
+  const mrr = perSubstring.reduce((s, p) => s + p.rr, 0) / total;
+  return { perSubstring, hitAtN, mrr, n };
+}
+
+function evaluateCase(kase, { count, ms, isError }, quality) {
   const warnings = [];
   if (isError) warnings.push('error result');
   if (ms > WARN_LATENCY_MS) warnings.push(`latency ${ms}ms > ${WARN_LATENCY_MS}ms`);
   if ((kase.expect === 'browse' || kase.expect === 'idlookup') && count === 0) {
     warnings.push('0 results for a browse/id-lookup case (expected data present)');
   }
+  if (quality) {
+    for (const p of quality.perSubstring) {
+      if (!p.hit) {
+        warnings.push(p.rank === null
+          ? `topNContains miss: "${p.needle}" not found in results`
+          : `topNContains miss: "${p.needle}" found at rank ${p.rank} > topN ${quality.n}`);
+      }
+    }
+  }
   const status = warnings.length ? 'WARN' : 'PASS';
   return { status, warnings };
 }
@@ -109,7 +149,14 @@ async function runCase(memoryModule, kase) {
   const ms = Date.now() - started;
   const text = textOfResult(result);
   const count = countEntryLines(text);
-  const evalResult = evaluateCase(kase, { count, ms, isError });
+  // expect stays a plain string for legacy cases ('results'/'browse'/'idlookup').
+  // New quality cases use an object: { kind?, topNContains: [...], topN? }.
+  const expectObj = kase.expect && typeof kase.expect === 'object' ? kase.expect : null;
+  const expectKind = expectObj ? expectObj.kind : kase.expect;
+  const topNContains = expectObj && Array.isArray(expectObj.topNContains) ? expectObj.topNContains : null;
+  const cutoffN = expectObj && Number.isInteger(expectObj.topN) ? expectObj.topN : 5;
+  const quality = topNContains ? scoreTopNContains(resultLines(text), topNContains, cutoffN) : null;
+  const evalResult = evaluateCase({ ...kase, expect: expectKind }, { count, ms, isError }, quality);
   return {
     id: kase.id,
     label: kase.label,
@@ -119,6 +166,7 @@ async function runCase(memoryModule, kase) {
     isError,
     errMsg,
     top3: topN(text, 3),
+    quality,
     status: evalResult.status,
     warnings: evalResult.warnings,
   };
@@ -128,6 +176,12 @@ function printCase(row) {
   process.stdout.write(`\n[${row.status}] ${row.id} — ${row.label}\n`);
   process.stdout.write(`  params: ${JSON.stringify(row.args)}\n`);
   process.stdout.write(`  results: ${row.count}  latency: ${row.ms}ms${row.isError ? `  ERROR: ${row.errMsg}` : ''}\n`);
+  if (row.quality) {
+    process.stdout.write(`  hit@${row.quality.n}: ${row.quality.hitAtN.toFixed(2)}  MRR: ${row.quality.mrr.toFixed(2)}\n`);
+    for (const p of row.quality.perSubstring) {
+      process.stdout.write(`    substr "${p.needle}" -> rank ${p.rank ?? 'none'}${p.hit ? '' : '  (miss)'}\n`);
+    }
+  }
   if (row.top3.length) {
     for (const line of row.top3) process.stdout.write(`    - ${line}\n`);
   } else {
@@ -138,18 +192,27 @@ function printCase(row) {
 
 function printSummary(rows) {
   process.stdout.write('\n=== recall-bench summary ===\n');
-  const widths = { id: 18, results: 8, ms: 8, status: 6 };
+  const widths = { id: 18, results: 8, ms: 8, status: 6, hit: 9, mrr: 7 };
   process.stdout.write(
-    `${'case'.padEnd(widths.id)}${'results'.padEnd(widths.results)}${'ms'.padEnd(widths.ms)}${'status'.padEnd(widths.status)}notes\n`,
+    `${'case'.padEnd(widths.id)}${'results'.padEnd(widths.results)}${'ms'.padEnd(widths.ms)}${'status'.padEnd(widths.status)}${'hit@N'.padEnd(widths.hit)}${'MRR'.padEnd(widths.mrr)}notes\n`,
   );
   for (const r of rows) {
+    const hitStr = r.quality ? r.quality.hitAtN.toFixed(2) : '-';
+    const mrrStr = r.quality ? r.quality.mrr.toFixed(2) : '-';
     process.stdout.write(
-      `${String(r.id).padEnd(widths.id)}${String(r.count).padEnd(widths.results)}${String(r.ms).padEnd(widths.ms)}${String(r.status).padEnd(widths.status)}${r.warnings.join('; ')}\n`,
+      `${String(r.id).padEnd(widths.id)}${String(r.count).padEnd(widths.results)}${String(r.ms).padEnd(widths.ms)}${String(r.status).padEnd(widths.status)}${hitStr.padEnd(widths.hit)}${mrrStr.padEnd(widths.mrr)}${r.warnings.join('; ')}\n`,
     );
   }
   const warnCount = rows.filter((r) => r.status === 'WARN').length;
   const totalMs = rows.reduce((s, r) => s + r.ms, 0);
-  process.stdout.write(`\ncases=${rows.length} pass=${rows.length - warnCount} warn=${warnCount} total_latency=${totalMs}ms\n`);
+  const qualityRows = rows.filter((r) => r.quality);
+  let aggLine = '';
+  if (qualityRows.length) {
+    const aggHit = qualityRows.reduce((s, r) => s + r.quality.hitAtN, 0) / qualityRows.length;
+    const aggMrr = qualityRows.reduce((s, r) => s + r.quality.mrr, 0) / qualityRows.length;
+    aggLine = `  agg_hit@N=${aggHit.toFixed(3)} agg_MRR=${aggMrr.toFixed(3)} (${qualityRows.length} scored cases)`;
+  }
+  process.stdout.write(`\ncases=${rows.length} pass=${rows.length - warnCount} warn=${warnCount} total_latency=${totalMs}ms${aggLine}\n`);
 }
 
 async function main() {

package/scripts/recall-quality-cases.json

@@ -0,0 +1,12 @@
+[
+  { "id": "q-kr-recall-phrase", "label": "Korean distinctive phrase: persistent memory", "args": { "query": "영속 메모리" }, "expect": { "kind": "results", "topNContains": ["영속 메모리"], "topN": 5 } },
+  { "id": "q-en-recall-pipeline", "label": "English distinctive phrase: recall fast-track builder", "args": { "query": "buildRecallFastTrackQuery" }, "expect": { "kind": "results", "topNContains": ["buildRecallFastTrackQuery"], "topN": 5 } },
+  { "id": "q-cycle1-drain", "label": "cycle1 ECONNRESET topic", "args": { "query": "cycle1 ECONNRESET" }, "expect": { "kind": "results", "topNContains": ["cycle1", "ECONNRESET"], "topN": 5 } },
+  { "id": "q-recall-general", "label": "bare recall keyword", "args": { "query": "recall" }, "expect": { "kind": "results", "topNContains": ["recall"], "topN": 5 } },
+  { "id": "q-uc4-topic-kr", "label": "UC4 topic query with Korean qualifier", "args": { "query": "recall period last 개선", "limit": 8 }, "expect": { "kind": "results", "topNContains": ["개선"], "topN": 5 } },
+  { "id": "q-recall-improve-kr", "label": "cache improvement topic", "args": { "query": "캐시 개선" }, "expect": { "kind": "results", "topNContains": ["개선"], "topN": 5 } },
+  { "id": "q-battle-balance", "label": "ProjectAA topic", "args": { "query": "ProjectAA" }, "expect": { "kind": "results", "topNContains": ["ProjectAA"], "topN": 5 } },
+  { "id": "q-fanout-recall-battle", "label": "fan-out array query cache+ProjectAA", "args": { "query": ["캐시 개선", "ProjectAA"], "limit": 5 }, "expect": { "kind": "results", "topNContains": ["개선", "ProjectAA"], "topN": 10 } },
+  { "id": "q-short-2tok-cycle", "label": "short 2-token cycle1 query", "args": { "query": "cycle1 drain", "limit": 8 }, "expect": { "kind": "results", "topNContains": ["cycle1"], "topN": 3 } },
+  { "id": "q-recall-scope-project", "label": "project-scoped recall query", "args": { "query": "recall", "cwd": "C:\\Project\\mixdog" , "limit": 10 }, "expect": { "kind": "results", "topNContains": ["recall"], "topN": 5 } }
+]

package/scripts/tool-smoke.mjs

@@ -1482,8 +1482,8 @@ if (!/repo-local|not web/i.test(codeGraphProps.mode?.description || '') || !/sou
 }
 const recallTool = MEMORY_TOOL_DEFS.find((tool) => tool.name === 'recall');
 const recallProps = recallTool?.inputSchema?.properties || {};
-if (!/when in doubt, recall first/i.test(recallTool?.description || '') || !recallProps.id?.anyOf || !/Do not invent ids/i.test(recallProps.id?.description || '')) {
-  throw new Error('recall schema must preserve prior-context guidance and id lookup shape');
+if (!/Use only to check prior conversation or past events/i.test(recallTool?.description || '') || !recallProps.id?.anyOf || !/Do not invent ids/i.test(recallProps.id?.description || '')) {
+  throw new Error('recall schema must preserve scoped prior-context guidance and id lookup shape');
 }
 if (!/array for independent fan-out/i.test(recallProps.query?.description || '') || !/Project pool selector/i.test(recallProps.projectScope?.description || '')) {
   throw new Error('recall schema must explain fan-out query and project scope filters');

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

@@ -8,13 +8,13 @@ Broad implementation agent.
 Bounded slices; smallest coherent change, not rewrite. Stop: unclear scope,
 growing blast radius, or Lead-only verification.
 
-EDIT-FIRST DISCIPLINE. Survey the slice with ONE batched read/grep round,
-then start patching the first bounded piece — broad scope means edit
-incrementally, not read exhaustively. NEVER "one more confirming read": if
-an anchor is plausible, the next call is `apply_patch`. Repeated read-only
-turns without an edit = stalling; any runtime reminder is your cue to patch
-the piece you understand or report blocked — a blocked report is a valid
-completion. Self-check comes AFTER edits; deep verification is Lead's.
+EDIT-FIRST DISCIPLINE. Survey the slice with ONE batched read/grep round, then
+patch the first bounded piece — broad scope means edit incrementally, not read
+exhaustively. NEVER "one more confirming read": if an anchor is plausible, the
+next call is `apply_patch`. Repeated read-only turns without an edit = stalling;
+any runtime reminder is your cue to patch the piece you understand or report
+blocked — a blocked report is a valid completion. Self-check comes AFTER edits;
+deep verification is Lead's.
 
 Minimal checks + how-to-verify. Hand off outcome as fragments anchored to
 `file:line`; no narration, no report bloat.

package/src/agents/scheduler-task/AGENT.md

@@ -7,6 +7,5 @@ toolSchemaProfile: read-write-search
 
 Scheduled task execution agent.
 
-Output a concise English handoff for Lead: outcome, action taken or skipped,
-notable evidence, and any remaining risk. If there is nothing actionable, use
-the skip protocol.
+Concise English handoff for Lead: outcome, action taken or skipped, notable
+evidence, remaining risk. Nothing actionable → use the skip protocol.

package/src/agents/webhook-handler/AGENT.md

@@ -7,6 +7,5 @@ toolSchemaProfile: read-write-search
 
 Inbound webhook event analysis agent.
 
-Output a concise English handoff for Lead: event outcome, action taken or
-skipped, notable evidence, and any remaining risk. If there is nothing
-actionable, use the skip protocol.
+Concise English handoff for Lead: event outcome, action taken or skipped,
+notable evidence, remaining risk. Nothing actionable → use the skip protocol.

package/src/agents/worker/AGENT.md

@@ -7,15 +7,14 @@ 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
-them 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;
-any runtime reminder is your cue to patch now or return blocked with what's
-missing — a blocked report is a valid completion. Confidence 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: 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; any runtime
+reminder is your cue to 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.
 
-Hand off outcome as fragments anchored to `file:line`; no narration, no
-report bloat.
+Hand off outcome as fragments anchored to `file:line`; no narration, no report
+bloat.
 

package/src/app.mjs

@@ -4,8 +4,12 @@ import { fileURLToPath } from 'node:url';
 import { performance } from 'node:perf_hooks';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
-const VALUE_OPTIONS = new Set(['--provider', '--model']);
+// --workflow is consumed (with its value) but ignored on the headless role
+// path: workflow selection is Lead-session-scoped. Accepting it keeps a
+// mistakenly passed `--workflow X` from leaking "X" into the role message.
+const VALUE_OPTIONS = new Set(['--provider', '--model', '--effort', '--workflow']);
 const FLAG_OPTIONS = new Set(['--readonly', '--help', '-h', '--plain', '--react', '--remote', '--onboarding']);
+const HEADLESS_FLAG_OPTIONS = new Set(['--fast']);
 const HEADLESS_ROLE_ALIASES = new Map([
   ['explorer', 'explore'],
   ['explore', 'explore'],
@@ -52,6 +56,7 @@ function unknownOption(argv) {
       continue;
     }
     if (FLAG_OPTIONS.has(arg)) continue;
+    if (HEADLESS_FLAG_OPTIONS.has(arg)) continue;
     if (String(arg || '').startsWith('-')) return arg;
   }
   return null;
@@ -66,6 +71,7 @@ function positionalArgs(argv) {
       continue;
     }
     if (FLAG_OPTIONS.has(arg)) continue;
+    if (HEADLESS_FLAG_OPTIONS.has(arg)) continue;
     if (String(arg || '').startsWith('-')) continue;
     out.push(arg);
   }
@@ -109,12 +115,15 @@ export async function run(argv = []) {
 
   const provIdx = argv.indexOf('--provider');
   const modelIdx = argv.indexOf('--model');
+  const effortIdx = argv.indexOf('--effort');
   const toolMode = argv.includes('--readonly') ? 'readonly' : 'full';
   const remote = argv.includes('--remote');
   const forceOnboarding = argv.includes('--onboarding');
   const opts = {
     provider: provIdx >= 0 ? argv[provIdx + 1] : undefined,
     model: modelIdx >= 0 ? argv[modelIdx + 1] : undefined,
+    effort: effortIdx >= 0 ? argv[effortIdx + 1] : undefined,
+    fast: argv.includes('--fast'),
     toolMode,
     remote,
     forceOnboarding,
@@ -150,6 +159,8 @@ export async function run(argv = []) {
       message: headless.message,
       provider: opts.provider,
       model: opts.model,
+      effort: opts.effort,
+      fast: opts.fast,
       cwd: process.cwd(),
     });
   }

package/src/headless-role.mjs

@@ -25,7 +25,7 @@ function makeTag(agent) {
     .replace(/^-+|-+$/g, '');
 }
 
-export function buildHeadlessSpawnArgs({ agent, tag, cwd, message, provider, model } = {}) {
+export function buildHeadlessSpawnArgs({ agent, tag, cwd, message, provider, model, effort, fast } = {}) {
   const spawnArgs = {
     type: 'spawn',
     agent: clean(agent),
@@ -35,6 +35,8 @@ export function buildHeadlessSpawnArgs({ agent, tag, cwd, message, provider, mod
   };
   if (clean(provider)) spawnArgs.provider = clean(provider);
   if (clean(model)) spawnArgs.model = clean(model);
+  if (clean(effort)) spawnArgs.effort = clean(effort);
+  if (fast === true) spawnArgs.fast = true;
   return spawnArgs;
 }
 
@@ -56,6 +58,8 @@ export async function runHeadlessRole({
   message,
   provider,
   model,
+  effort,
+  fast,
   cwd = process.cwd(),
   write = (text) => stdout.write(text),
   writeErr = (text) => stderr.write(text),
@@ -86,6 +90,8 @@ export async function runHeadlessRole({
     message: cleanMessage,
     provider,
     model,
+    effort,
+    fast,
   });
 
   let taskId = null;

package/src/mixdog-session-runtime.mjs

@@ -54,13 +54,12 @@ import {
 } from './runtime/agent/orchestrator/providers/custom-tool-wire.mjs';
 import {
   channelSetup,
-  deleteChannel,
   deleteSchedule,
   deleteWebhook,
   forgetDiscordToken,
   forgetTelegramToken,
   forgetWebhookAuthtoken,
-  saveChannel,
+  setChannel,
   saveDiscordToken,
   saveTelegramToken,
   saveSchedule,
@@ -138,6 +137,7 @@ import {
   normalizeAutoClearConfig,
   resolveAutoClearIdleMs,
   autoClearIdleMsForProvider,
+  autoClearProviderDefaults,
   normalizeCompactionConfig,
   moduleEnabled,
   setModuleEnabledInConfig,
@@ -177,6 +177,7 @@ import {
   agentPresetSlot,
   normalizeAgentId,
   normalizeWorkflowId,
+  DEFAULT_WORKFLOW_ID,
   createWorkflowHelpers,
   normalizeSearchProviderId,
   isDefaultSearchRouteConfig,
@@ -638,7 +639,7 @@ export async function createMixdogSessionRuntime({
   let updateProcessState = { phase: 'idle', version: null, error: null };
 
   function autoUpdateEnabled() {
-    return config?.update?.auto === true;
+    return config?.update?.auto !== false;
   }
 
   async function checkForUpdateInternal({ force = false } = {}) {
@@ -681,8 +682,8 @@ export async function createMixdogSessionRuntime({
   // Non-blocking boot hook: fires after the runtime object below is fully
   // constructed (setTimeout(0) defers past the synchronous return), so a
   // slow/hanging registry request or npm install can never delay session
-  // boot. Auto-update defaults to OFF (config.update.auto is undefined until
-  // the user opts in via setAutoUpdate(true)); a failed check or install is
+  // boot. Auto-update defaults to ON unless config.update.auto is explicitly
+  // false; a failed check or install is
   // swallowed — getUpdateStatus()/getUpdateSettings() are the only surfaces,
   // there is no push notice channel from runtime -> TUI today.
   // force:true — always hit the registry at boot (the 24h disk cache went
@@ -1662,6 +1663,7 @@ export async function createMixdogSessionRuntime({
     hasOwn,
     normalizeAutoClearConfig,
     autoClearIdleMsForProvider,
+    autoClearProviderDefaults,
     normalizeCompactionConfig,
     normalizeCompactTypeSetting,
     normalizeSystemShellConfig,
@@ -1749,10 +1751,14 @@ export async function createMixdogSessionRuntime({
     },
     get workflow() {
       const dataDir = cfgMod.getPluginData?.() || STANDALONE_DATA_DIR;
+      const active = activeWorkflowSummary(config, dataDir);
       if (session?.workflow && typeof session.workflow === 'object') {
-        return workflowSummary(session.workflow);
+        const current = workflowSummary(session.workflow);
+        return current?.id && active?.id && current.id !== active.id
+          ? { ...active, currentSession: current, appliedToCurrentSession: false }
+          : active;
       }
-      return activeWorkflowSummary(config, dataDir);
+      return active;
     },
     get outputStyle() {
       return getOutputStyleStatusCached().current;
@@ -2003,13 +2009,8 @@ export async function createMixdogSessionRuntime({
       remoteStateListeners.add(listener);
       return () => remoteStateListeners.delete(listener);
     },
-    saveChannel(entry) {
-      const result = saveChannel(entry);
-      reloadChannelsSoon();
-      return result;
-    },
-    deleteChannel(name) {
-      const result = deleteChannel(name);
+    setChannel(entry) {
+      const result = setChannel(entry);
       reloadChannelsSoon();
       return result;
     },

package/src/output-styles/default.md

@@ -8,24 +8,21 @@ keep-coding-instructions: true
 # Output Style
 
 Mixdog default — the most detailed style, but always summary-form, never
-essay-form. Depth comes from picking the right facts, not from explaining more.
+essay-form. Depth comes from picking the right facts, not explaining more.
 
 Content
 - Lead with the outcome in one short sentence, then only the detail that
   matters: what changed, key evidence (paths, commands, errors, verification).
-- Summarize at the concept level: name the problem/behavior and the direction,
-  not the code path. Cite a symbol or path only as an anchor, never as the
-  explanation.
+- Summarize at the concept level: name the problem/behavior and direction, not
+  the code path. Cite a symbol/path only as an anchor, never as the explanation.
 - Compress by cutting content (filler, hedging, connective padding, restated
   facts), not by clipping grammar: keep natural, complete sentences in the
   user's language — never telegraph-style stub endings. Technical terms and
   code stay exact.
 - State conclusions, not reasoning: no mechanism walkthroughs, background, or
-  chained qualifiers unless the user asks. One decisive fact beats three
-  hedges.
-- Say each point once: problem and fix in ONE compact statement, not a
-  restated pair. Prefer fewer, denser items over full coverage of every
-  nuance.
+  chained qualifiers unless asked. One decisive fact beats three hedges.
+- Say each point once: problem and fix in ONE compact statement, not a restated
+  pair. Prefer fewer, denser items over covering every nuance.
 - Size budget: roughly TWICE the Simple style — per point about 2 rendered
   lines, whole report ~10–15 lines. Spend the extra room on evidence and
   context Simple would drop, not on longer sentences.
@@ -39,14 +36,13 @@ Content
 - Keep paths, commands, symbols, API names, code, and exact errors verbatim.
 
 Layout (hard rules)
-- One bullet or numbered item = one idea, at most 2 rendered lines including
-  its sub-bullet. If it needs more, cut the detail — do not add lines.
-- Open each list item with a short **bold key point**, then the brief
-  elaboration — never bury the point mid-sentence.
-- Insert a blank line between numbered items, and between any list items that
-  run past one line (loose list). Never emit a wall of consecutive multi-line
-  items.
-- Keep paragraphs to ~3 lines max, with a blank line between paragraphs,
-  lists, and code blocks.
+- One bullet or numbered item = one idea, at most 2 rendered lines including its
+  sub-bullet. If it needs more, cut the detail — do not add lines.
+- Open each item with a short **bold key point**, then the brief elaboration —
+  never bury the point mid-sentence.
+- Insert a blank line between numbered items, and between any list items running
+  past one line (loose list). Never emit a wall of consecutive multi-line items.
+- Keep paragraphs to ~3 lines max, with a blank line between paragraphs, lists,
+  and code blocks.
 - Nest at most one sub-level; deeper detail means you are over-explaining.
 - Never name this style unless asked.

package/src/output-styles/minimal.md

@@ -11,16 +11,15 @@ Minimal — a very short summary: one or two sentences, nothing more.
 
 - Summarize only the net result in one short sentence; add a second short
   sentence only if a second fact (verification, blocker) genuinely needs it.
-  Never cram unrelated facts into one run-on line just to stay at one sentence.
+  Never cram unrelated facts into a run-on just to stay at one sentence.
 - Size budget: roughly HALF the Simple style — 1–2 plain sentences, ~2–3
   rendered lines at most, however large the task was.
 - Compress by cutting content, not grammar: natural, complete sentences only.
   Concept-level only — never walk through code or mechanisms.
-- Summarize, never itemize: do not describe which files changed or how they were
-  edited. State only what the change accomplishes.
+- Summarize, never itemize: do not describe which files changed or how. State
+  only what the change accomplishes.
 - No headings, bullets, numbered lists, labels, or sections — plain sentences
-  only. This holds even when the request says "report" or "summary"; keep it to
-  one or two sentences regardless.
+  only, even when the request says "report" or "summary".
 - Preferred pattern: `<target> 변경되었습니다. <verification> 통과 완료입니다.`
 - If verification was not run, say the change is done and verification was not
   run.

package/src/output-styles/simple.md

@@ -12,8 +12,8 @@ Practical concise — outcome-first handoffs for coding work: summarize the
 result, do not narrate or explain the change.
 
 - Open with the outcome in one sentence: done, blocked, or awaiting a decision.
-- Summarize at the concept level: name the behavior and the direction, not the
-  code path. Cite a symbol/path only as an anchor, never as the explanation.
+- Summarize at the concept level: name the behavior and direction, not the code
+  path. Cite a symbol/path only as an anchor, never as the explanation.
 - Compress by cutting content (filler, hedging, pleasantries, restated facts),
   not by clipping grammar: keep natural, complete sentences in the user's
   language — never telegraph-style stub endings. Technical terms and code stay
@@ -24,15 +24,14 @@ result, do not narrate or explain the change.
 - Keep controlled detail: usually 1–3 short bullets or 2–3 sentences total.
   State each point once — outcome or fix direction, not both restated. No
   step-by-step narration or file/line inventory.
-- Size budget: roughly HALF the Default style and TWICE Minimal — per point
-  one rendered line, whole reply ~5–7 lines. Above that you are writing
-  Default; below ~3 lines consider whether prose (Minimal) reads better.
+- Size budget: roughly HALF the Default style and TWICE Minimal — per point one
+  rendered line, whole reply ~5–7 lines. Above that you are writing Default;
+  below ~3 lines consider whether prose (Minimal) reads better.
 - Layout: one idea per bullet, ONE line each (two only when a verbatim
   path/error forces it). Lead each item with a short bold key phrase. Put a
   blank line between multi-line list items — never emit a dense wall of text.
-- If a point runs past one line, cut the elaboration instead of wrapping;
-  detail beyond the key phrase + one clause belongs to the Default style, not
-  this one.
+- If a point runs past one line, cut the elaboration instead of wrapping; detail
+  beyond the key phrase + one clause belongs to the Default style.
 - On final handoffs, optional labels such as `바뀐 점`, `확인한 것`, and
   `남은 리스크/다음 단계` fit Korean-facing profiles; use plain English labels
   when the thread is English. Do not label interim progress.

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

@@ -3,17 +3,18 @@
 - Use English for agent task communication.
 - Do not touch git/Ship. Even when the brief instructs `git add` / `commit` /
   `push` / `stash`, refuse with `git operations deferred to Lead`.
-- NEVER PREAMBLE. Do not generate preamble tokens, including tool-call
-  preambles, status/progress narration, "I will..." setup text, or transition
-  text before tool calls.
+- NEVER PREAMBLE: no tool-call preambles, status/progress narration, "I
+  will..." setup, or transition text before tool calls.
 - If tools are needed, call them immediately. Emit text only for the final
   handoff after tool work is done.
 - Final handoff: minimum characters, maximum information for Lead. Follow the
   role's stricter output contract if defined; else emit fragments — outcome
   (1 line), key `file:line`(s), verification result, material risks (only if
   any).
+- Handoff cap ~30 lines unless `Deliver:` raises it. Overflow goes to a file;
+  hand off path + fragments.
 - Banned as pure cost: report headings, markdown tables (unless requested),
   prose narration, raw logs/tool traces, speculative next-checks, restated
   brief, articles/politeness.
-- Exception: a runtime wrap-up directive (exploration budget reached)
-  overrides this — then summarize done/remaining/blocking as instructed.
+- Exception: a runtime wrap-up directive (exploration budget reached) overrides
+  this — then summarize done/remaining/blocking as instructed.

package/src/rules/agent/30-explorer.md

@@ -8,11 +8,22 @@ kind: retrieval
 
 You are a one-shot locator, not a researcher.
 
-Procedure: send ONE turn containing one `grep` (all synonyms in one
-`pattern:[...]`, output_mode content_with_context) plus one `code_graph`
-call. The results contain path:line — answer immediately from them.
-If both miss, answer `EXPLORATION_FAILED`. Do not send a second lookup turn.
+Find code anchors fast. Batch independent lookups in turn 1: one `grep`
+with all literal strings/errors/identifier guesses, plus `code_graph`/`find`
+or `glob` when useful.
+
+Prefer user-visible strings and identifiers. `dist`/generated hits are leads,
+not answers: trace them back to source. On miss, derive new tokens from
+returned paths/names or widen scope; do not repeat synonyms. NEVER issue
+serial single-tool turns: every turn after the first must batch ALL remaining
+candidate lookups in one shot.
+
+Budget: 4 tool turns. The FIRST credible `path:line` ends the search — answer
+with what you have; more anchors are not more value. Spend remaining turns
+only while you have zero anchors, and buy coverage with wider batches, never
+with extra serial turns. On ANY iteration warning, answer immediately with
+best anchors or `EXPLORATION_FAILED`.
 
 Answer format, nothing else:
-- up to 5 lines of `path:line — symbol/name — short reason` (append `?` if weak)
+- up to 5 lines: `path:line — symbol/name — short reason` (append `?` if weak)
 - or `EXPLORATION_FAILED`

package/src/rules/lead/01-general.md

@@ -1,11 +1,11 @@
 # General
 
 - Omit direct names, honorifics, headings, and labels in preambles.
-- Preambles are optional: use them only when they add user-visible value, keep
-  them to one short sentence, and skip routine lookup narration.
+- Preambles are optional: use only when they add user-visible value, keep to
+  one short sentence, skip routine lookup narration.
 - Destructive/hard-to-reverse actions require explicit confirmation.
 - Never push, build, or deploy without an explicit user request.
   Implementation approval is not deploy approval.
-- Rather than deferring checks or work to the user, proactively handle whatever
-  you can first, and propose only the parts that need a decision in a
-  consultative tone ("Shall we proceed this way?").
+- Proactively handle what you can first rather than deferring to the user;
+  propose only the parts needing a decision, consultatively ("Shall we
+  proceed this way?").

package/src/rules/lead/lead-brief.md

@@ -3,9 +3,12 @@
 - Brief = one-line fragments `Goal:` `Anchors:` `Allow/Forbid:` `Deliver:`
   `Verify:` (+`Stop:` heavy-worker). No role-known rules, background, or
   motivation — minimum characters, maximum information.
-- Full brief only on fresh spawn or `respawned: true` (dead-tag send =
-  cold session; re-supply anchors). Live-session follow-ups = delta only;
-  never restate Goal/rules.
+- `Deliver:` states output size/shape (e.g. "fragments <=15 lines", "verdict
+  + top-3 risks", "detail to file, path only"). Never request a long report
+  in the handoff itself.
+- Full brief only on fresh spawn or `respawned: true` (dead-tag send = cold
+  session; re-supply anchors). Live-session follow-ups = delta only; never
+  restate Goal/rules.
 - Never `send` mid-run; batch all adjustments into ONE follow-up after
   completion. Interrupt only to cancel.
 - All agent communication in English.

package/src/rules/lead/lead-tool.md

@@ -1,6 +1,8 @@
 # Lead Tool Use
 
-- Lead owns repo-local shell work: run git/build/test/verification commands via
-  `shell` directly; do not delegate them to agents.
-- Use the current project/workspace selected by the session. Only change the work project when the user asks for a different project or a tool call explicitly needs another project root.
-- Use `agent` for scoped implementation, research, review, and debugging, not for git commit/push/stash or Ship.
+- Lead owns repo-local shell work: run git/build/test/verification via `shell`
+  directly; do not delegate to agents.
+- Use the session's current project/workspace. Change the work project only
+  when the user asks for another project or a tool call needs another root.
+- Use `agent` for scoped implementation, research, review, and debugging — not
+  for git commit/push/stash or Ship.