agentlas 1.0.22 → 1.0.23

package/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## 1.0.23 — 2026-08-01
+
+- **Semantic routing stays with the connected model.** Image-capability and
+  installed-agent routing no longer retain regex hints, keyword glossaries,
+  deterministic role vetoes, or caller-authored fallback labels. If the model
+  cannot return a valid judgment, the decision remains unavailable.
+- **Fresh Terminal databases match Desktop schema 86.** The retired
+  chat-level hired-agent roster is absent from the bootstrap schema; project
+  order and current-turn task-force targets own controller and helper binding.
+
 ## 1.0.22 — 2026-08-01
 
 - **Task classes no longer come from a fixed keyword dictionary.** An explicit

package/README.md

@@ -54,33 +54,21 @@ agentlas
 agentlas "Identify why repository tests are failing"      # Uses the connected project's first agent as controller
 agentlas run -p "Summarize the last 3 CHANGELOG entries"  # Prints the project controller's final answer to stdout
 git log --oneline -20 | agentlas run                      # Reads stdin; still requires a connected project
-agentlas run <agent> "Advanced direct request"            # Exact-agent direct invocation; no project substitution
-agentlas <agent>                                          # Starts an interactive REPL with an exact agent
+agentlas run <agent> "Advanced direct request"            # Explicit one-turn exact-agent override
 agentlas firm <firm> "Request"                            # Delegates to an exact firm CEO
 ```
 
 Ordinary one-shot work is project-first: run it inside a folder connected to a Desktop Work project. The first agent in that project's ordered pool owns the task; remaining members are eligible task-scoped sub-agents. Missing projects, empty teams, and unavailable controllers stop without substituting another agent. The separate `agentlas route` preview is model-judged and stays unresolved when no exact installed agent can be justified.
 
-**3. Resuming & Managing Conversations**
-
-```sh
-agentlas chats            # List recent conversations (shares SQLite DB with Desktop app)
-agentlas chats 30         # List last 30 conversations
-agentlas open <chat-id>   # Re-open REPL with the chat's assigned agent
-```
-
 ## Command Reference
 
 Run `agentlas help` for the authoritative command list. Below is the organized command structure:
 
-### TALK & RUN
+### PROJECT WORK
 ```sh
-agentlas <agent>                 # Chat with agent (equivalent to chat <agent>)
-agentlas chat <agent>
-agentlas run [agent] [prompt]    # Single-run execution (-p, --runtime, --permission, stdin)
+agentlas                         # Open the connected project's controller session
+agentlas run [agent] [prompt]    # Project-first run; optional agent is one explicit turn
 agentlas firm <firm> [task]      # Delegate task to firm CEO
-agentlas chats [n]               # List recent chat conversations
-agentlas open <chat-id>          # Resume conversation by ID
 agentlas cd <agent>              # Print folder path of an agent (cd "$(agentlas cd x)")
 ```
 
@@ -164,7 +152,7 @@ $ agentlas lst
 Command list: agentlas help  ·  To run this exact prompt: agentlas run -p "lst"
 ```
 
-Desktop-only surface names (`site`, `trex`, `prompts`, `dashboard`, `marketplace`, `library`, `groups`, `settings`, `apps`, `quests`, `bookmarks`, `one`) also halt with typo guard guidance directing you to equivalent terminal commands. To run a single word as a prompt, wrap it in quotes or pass `run -p`.
+Desktop-only surface names (`site`, `trex`, `prompts`, `dashboard`, `marketplace`, `library`, `settings`, `apps`, `quests`, `bookmarks`, `one`) also halt with typo guard guidance directing you to equivalent terminal commands. To run a single word as a prompt, wrap it in quotes or pass `run -p`.
 
 ### Permission Levels
 
@@ -176,9 +164,9 @@ Desktop-only surface names (`site`, `trex`, `prompts`, `dashboard`, `marketplace
 
 *Note: Saved `full` permission is session-bound and automatically fails closed to `write` on subsequent executions. REPL shell execution (`!<command>`) runs strictly under `full` permission, enforced with 8MB output capping, secret masking, and process-group termination.*
 
-## REPL & Orca Multi-Session Architecture
+## Project Work Execution Tree
 
-Executing `agentlas` with no arguments starts an interactive REPL session. Foreground execution turns run inside an active session, while subagents created via `/spawn` execute concurrently in background sessions. The terminal streams output from the currently focused active session and displays one-line notices when background tasks finish.
+Executing `agentlas` with no arguments resolves the current folder to an Agentlas project and starts its first ordered agent as the task controller. The controller may create task-scoped worker runs; users can inspect the real execution tree, but cannot replace the controller or address workers around it.
 
 ### REPL Slash Commands
 
@@ -187,14 +175,9 @@ Executing `agentlas` with no arguments starts an interactive REPL session. Foreg
 | `/help` | Display available commands and keybindings |
 | `/sessions` · `/tree` | View session table or parent-child process tree |
 | `/s <n>` · `/switch <n>` | Switch active session (replays tail logs & subscribes to live stream) |
-| `/spawn <agent> [task]` | Spawn subagent background session (executes task immediately if supplied) |
-| `/steer <n> <msg>` | Queue instruction for session `n`'s next turn |
 | `/kill <n>` | Interrupt active turn in session `n` |
 | `/rm <n>` | Remove session `n` |
-| `/broadcast <msg>` | Broadcast instruction to all running sessions |
-| `/use <agent>` | Change active agent in main session |
 | `/agents` · `/list` | List installed agents |
-| `/chats [n]` | View recent conversations |
 | `/mcp` | List active MCP servers |
 | `/doctor` | Run runtime & database diagnostics |
 | `/runtime <kind>` | Set runtime for new sessions (`claude-code` \| `codex` \| `gemini`) |
@@ -242,9 +225,9 @@ The launcher (`bin/agentlas.cjs`) runs system Node against `engine/`. The defaul
 | Windows | `%APPDATA%\Agentlas` |
 | Linux | `$XDG_CONFIG_HOME/Agentlas` (default `~/.config/Agentlas`) |
 
-The SQLite database file is `agentlas.sqlite` (`user_version=85`). When launched for the first time without an existing database, it bootstraps schemas using `engine/bootstrap-schema.sql`. Consequently, **projects, installed agents, task history, automation sessions, and MCP registrations are shared across Desktop and Terminal**. Subagent sessions spawned via `/spawn` are stored as scoped execution ledgers rather than new global Work conversations.
+The SQLite database file is `agentlas.sqlite` (`user_version=86`). When launched for the first time without an existing database, it bootstraps schemas using `engine/bootstrap-schema.sql`. Consequently, **projects, installed agents, task history, automation sessions, and MCP registrations are shared across Desktop and Terminal**. The first ordered project agent remains the controller; additional agents are task-scoped and are stored only as execution ledgers, never as global conversations or durable owners.
 
-SQLite Driver Fallback Ladder: `better-sqlite3` (optionalDependency native build) -> Node 22+ `node:sqlite`.
+SQLite driver priority: `better-sqlite3` (optional dependency native build), then Node 22+ `node:sqlite` when the first driver is unavailable.
 
 ### Hub Installation & Safety Policy
 

package/engine/agentlas-banner.cjs

@@ -195,7 +195,6 @@ function renderBanner(ctx) {
   const menu = [
     [ui.t("banner.menu.help"), "/help"],
     [ui.t("banner.menu.sessions"), "/sessions"],
-    [ui.t("banner.menu.chats"), "/chats"],
     [ui.t("banner.menu.quit"), "/quit"],
   ];
   const infoRoom = inner - CARD_PAD * 2;

package/engine/agentlas-capabilities.cjs

@@ -40,58 +40,9 @@ function runtimeFromSpec(spec) {
   return CLI_KINDS.includes(spec) ? { mode: "cli", kind: spec } : { mode: "api", backend: spec, model: null };
 }
 
-// Does this agent's job involve generating/handling images?
-// /그림(?!자)/ — "그림자"(shadow)는 이미지 힌트가 아니다.
-const IMAGE_HINTS = [
-  /image/i, /이미지/, /그림(?!자)/, /\bdesign\b/i, /디자인/, /쇼핑몰/, /상품\s*(사진|이미지|상세)/, /상세\s*페이지/,
-  /thumbnail/i, /썸네일/, /banner/i, /배너/, /poster/i, /포스터/, /visual/i, /비주얼/, /illustrat/i, /일러스트/,
-  /로고/, /\blogo\b/i, /사진/, /photo/i, /nano-?banana/i, /imagen/i, /이미지\s*생성/, /그래픽/, /graphic/i,
-];
-// 빌더/메타/조율/거버넌스 역할은 (이미지 에이전트를 *만들* 수는 있어도) 스스로 이미지를 생산하지 않는다.
-// 이런 역할이 system_prompt에 "이미지/디자인"을 언급한다는 이유로 gemini로 끌려가면 코드/빌드 품질이 떨어진다.
-const NON_IMAGE_ROLES = new Set(["meta", "builder", "orchestrator", "pm", "curator", "governance"]);
-// 부정/거절 문장("이미지 생성 금지", "영상·이미지 생성은 하지 않는다")은 능력이 아니라
-// 반(反)능력 선언이다 — 그 문장 안의 힌트는 세지 않는다. 단, "묻지 않고 바로 생성한다"처럼
-// 긍정문에 흔한 보조 부정("않고","없다" 단독)은 잡지 않도록 강한 금지 구문만 매칭한다.
-const IMAGE_NEGATION_RE = /(금지|하지\s*않|하지\s*마|말\s*것|거절|불가(?!피)|아니다|refuse|\bnever\b|\bdo(es)?\s+not\b|\bdon'?t\b)/i;
-// 이미지 생성 도구 이름은 단독으로도 이미지 생산 역할의 강한 증거다.
-const IMAGE_TOOL_MARKERS = [/nano-?banana/i, /\bimagen\b/i, /gpt-image/i, /grok\s*imagine/i];
-// 사고(2026-07-12): appbridge CEO 프롬프트의 "코드/디자인/스토어 결정의 owner가 아니다" 속
-// "디자인" 한 단어로 이미지 에이전트 판정 → PPT 요청 세션이 통째로 gemini로 전환됐다.
-// 수리: 정체성 존(이름/태그라인)은 그대로 신뢰하되, 본문 단독으로는 "힌트를 포함한 긍정문
-// 3문장 이상"을 요구한다. 문장 단위로 세므로 "상품 이미지 생성 금지" 같은 한 문장이
-// 겹치는 정규식 여러 개를 동시에 때려도 1클러스터다.
-const MIN_BODY_IMAGE_SENTENCES = 3;
-const BODY_SCAN_CAP = 16000; // 로컬 임포트 상한과 동일 — 클라우드 무제한 프롬프트의 전문 스캔 방지
-// 어휘 스코어 — IMAGE_HINTS 단어목록에 대한 참고용 스코어러. 하우스 룰(2026-07-25):
-// 단어목록은 최종 결정을 내리지 않는다. resolveNeedsImage가 IMAGE_HINTS를 힌트로만 모델에
-// 넘기고, 연결 모델이 없으면 이미지 여부를 판정하지 않는다(어휘 폴백 제거). 이 함수는 힌트
-// 품질을 고정하는 참조 스코어러로만 남는다(회귀 테스트가 정밀도를 검증).
-function needsImageLexical(agent) {
-  if (!agent) return false;
-  if (NON_IMAGE_ROLES.has(String(agent.role || "").toLowerCase())) return false;
-  // 정체성 존은 사용자가 선언한 이름/태그라인만 — slug는 폴더명에서 기계 파생되므로
-  // ("design-system" 리포 임포트 등) 단독 신뢰 대상이 아니다.
-  const identity = [agent.name, agent.name_en, agent.tagline, agent.tagline_en].filter(Boolean).join(" ");
-  if (IMAGE_HINTS.some((re) => re.test(identity))) return true;
-  // 팀 CEO 두뇌(부서 소개·위임 규칙)에서는 body 키워드가 역할 증거가 아니다 — vibecoder처럼
-  // "Design and Publishing HQ" 부서명이 10문장씩 나오는 조율용 프롬프트가 이미지 팀으로
-  // 오판되던 사례. entity_kind='team'은 사용자가 선언한 정체성 존만 신뢰한다.
-  if (String(agent.entity_kind || "").toLowerCase() === "team") return false;
-  const body = String(agent.system_prompt || "").slice(0, BODY_SCAN_CAP);
-  let clusters = 0;
-  for (const sentence of body.split(/[\n.!?。！？]+/)) {
-    if (!sentence || IMAGE_NEGATION_RE.test(sentence)) continue;
-    if (IMAGE_TOOL_MARKERS.some((re) => re.test(sentence))) return true;
-    if (IMAGE_HINTS.some((re) => re.test(sentence))) {
-      clusters += 1;
-      if (clusters >= MIN_BODY_IMAGE_SENTENCES) return true;
-    }
-  }
-  return false;
-}
-
-// ── 상주 판정 서비스 배선 — 모델이 의미로 최종 판정, IMAGE_HINTS는 참고 힌트 ──────
+// Image capability is judged only by the connected model from the complete
+// agent identity and instructions. No regex, keyword glossary, role veto, or
+// default runtime may make this semantic decision.
 // needsImage 호출자(REPL 배지·autoRuntimeFor·routingNote)는 동기라서 warm-cache 패턴:
 // 비동기 경로(resolveNeedsImage)가 먼저 판정해 캐시를 데우고, 동기 needsImage는 캐시만
 // 읽는다. 캐시 미스 = "이미지 아님"으로 처리(어휘 폴백 없음) — 모델 판정만 이미지로 인정한다.
@@ -101,22 +52,13 @@ function imageJudgeInput(agent) {
   const identity = [agent.slug, agent.name, agent.name_en, agent.tagline, agent.tagline_en].filter(Boolean).join(" | ");
   return `${identity}\n---\n${String(agent.system_prompt || "").slice(0, 6000)}`;
 }
-// 역할/팀 베토는 보수적 하드 가드로 유지 — 조율 두뇌가 부서 소개 문장("Design HQ")으로
-// 이미지 팀이 되던 사고(vibecoder/appbridge)는 모델 판정 대상에서 아예 뺀다.
-function imageJudgeVetoed(agent) {
-  if (!agent) return true;
-  if (NON_IMAGE_ROLES.has(String(agent.role || "").toLowerCase())) return true;
-  if (String(agent.entity_kind || "").toLowerCase() === "team") return true;
-  return false;
-}
 // 이 에이전트의 직무가 이미지 생산인지를 연결 모델이 의미로 판정한다.
 // 하우스 룰(2026-07-25): 연결 모델이 없으면 단어목록으로 이미지 여부를 결정하지 않는다.
 // 러너 없음/타임아웃/정크 → source:"unavailable"(판정 불가)로 반환하고, 호출자는 안전
 // 기본값(세션 런타임 유지, gemini/codex 하이재킹 금지)을 지킨다. 이미지 판정은 저위험
 // 능력 추론이라 실패-닫힘이 아니라 "판정 안 함"으로 정직하게 둔다.
 async function resolveNeedsImage(agent) {
-  // 팀/역할 하드 가드는 구조적 판단(키워드 아님) — 판정 대상에서 제외하고 결정적으로 not-image.
-  if (imageJudgeVetoed(agent)) return { image: false, source: "deterministic" };
+  if (!agent) return { image: false, source: "unavailable", decided: false };
   let judgment;
   try {
     judgment = require("./agentlas-judgment.cjs");
@@ -134,13 +76,11 @@ async function resolveNeedsImage(agent) {
     labels: ["image", "not-image"],
     multi: false,
     input,
-    hints: { image: IMAGE_HINTS.map((re) => re.source) },
     guidance:
       "Judge the agent's role from its identity and instructions, in any language. Mentioning images is not " +
       "producing them: builders, orchestrators, PMs, curators, and coordination brains that commission or " +
       "delegate image work are 'not-image'. Refusals or prohibitions ('never generate images') declare the " +
       "opposite of a capability.",
-    fallback: [],
   });
   // 모델이 판정을 못 냈다 → 어휘 폴백 없이 "판정 불가". 캐시하지 않아 이후 모델 연결 시 재판정한다.
   if (verdict.source !== "llm" || !verdict.labels.length) return { image: false, source: "unavailable", decided: false };
@@ -155,18 +95,16 @@ async function resolveNeedsImage(agent) {
 // Does this agent's job involve generating/handling images? Sync surface for badges and
 // autoRuntimeFor. House rule (2026-07-25): only a warm MODEL verdict counts as "image".
 // A cache miss (no model judgment / not warmed) is treated as "not specifically an image
-// agent" so a keyword guess can never hijack the runtime to gemini/codex. The lexical
-// scorer (needsImageLexical) survives only as the hint source for the judge, not a decision.
+// agent" so an unproven semantic guess can never hijack the runtime to gemini/codex.
 function needsImage(agent) {
   if (!agent) return false;
-  if (imageJudgeVetoed(agent)) return false;
   const cached = imageVerdicts.get(imageJudgeInput(agent));
   return cached ? cached.image : false;
 }
-// 라벨용 — 이 에이전트의 현재 이미지 판정이 모델("llm")인지 결정적 경로("deterministic")인지.
+// 라벨용 — 모델 판정이 없으면 unavailable이며 다른 의미 판정으로 대체하지 않는다.
 function imageJudgmentSource(agent) {
-  if (agent && !imageJudgeVetoed(agent) && imageVerdicts.has(imageJudgeInput(agent))) return "llm";
-  return "deterministic";
+  if (agent && imageVerdicts.has(imageJudgeInput(agent))) return "llm";
+  return "unavailable";
 }
 function clearImageJudgments() {
   imageVerdicts.clear();
@@ -195,7 +133,6 @@ module.exports = {
   specOf,
   runtimeFromSpec,
   needsImage,
-  needsImageLexical,
   resolveNeedsImage,
   imageJudgmentSource,
   clearImageJudgments,

package/engine/agentlas-core-harness.cjs

@@ -40,6 +40,8 @@ function runtimeRoots(explicitRoot) {
     binRoot,
     path.join(os.homedir(), ".agentlas", "runtime", "current"),
   ];
+  // Monorepo/development parity with Desktop's signed-bundle candidate.
+  roots.push(path.resolve(__dirname, "..", "..", "agentlas_desktop", "Hephaestus"));
   if (process.resourcesPath) roots.push(path.join(process.resourcesPath, "Hephaestus"));
   if (process.platform === "darwin") roots.push("/Applications/Agentlas.app/Contents/Resources/Hephaestus");
   return unique(roots);
@@ -77,13 +79,12 @@ function resolveCoreRuntimeRootFromCandidates(candidateRoots, requiredMarkers =
   const markers = [...CORE_RUNTIME_MARKERS, ...requiredMarkers];
   const minVersion = options.minVersion ? normalizeSemVer(String(options.minVersion)) : null;
   if (options.minVersion && !minVersion) throw new Error(`Invalid minimum Core version: ${options.minVersion}`);
+  const eligible = [];
   for (const root of unique(candidateRoots)) {
     try {
       if (!markers.every((segments) => fs.existsSync(path.join(root, ...segments)))) continue;
-      if (minVersion) {
-        const version = readCoreRuntimeVersion(root);
-        if (!version || compareSemVer(version, minVersion) < 0) continue;
-      }
+      const version = readCoreRuntimeVersion(root);
+      if (minVersion && (!version || compareSemVer(version, minVersion) < 0)) continue;
       // `~/.agentlas/runtime/current` 는 업데이터가 원자적으로 갈아 끼우는 심볼릭
       // 링크다. 그 경로를 그대로 넘기면 Python 이 import 를 늦게 해석하는 특성상,
       // 긴 실행 도중 업데이트가 일어나면 **옛 버전 모듈과 새 버전 모듈이 한 프로세스에
@@ -92,17 +93,23 @@ function resolveCoreRuntimeRootFromCandidates(candidateRoots, requiredMarkers =
       //
       // 실경로로 고정해도 라이브 버전 선택은 그대로다: **다음** 호출이 다시 해석해
       // 새 릴리스를 집는다. 없어지는 것은 실행 중 교체뿐이다.
-      try {
-        return fs.realpathSync(root);
-      } catch {
-        // 링크를 읽을 수 없다고 실행을 거부할 이유는 없다 — 이전 동작대로 경로를 쓴다.
-        return root;
-      }
+      let pinned = root;
+      try { pinned = fs.realpathSync(root); } catch {}
+      eligible.push({ root: pinned, version });
     } catch {
       // Continue to the next installed/bundled root.
     }
   }
-  return null;
+  // Desktop and Terminal must attach to the same newest valid Core. Candidate
+  // order is only a tie-breaker; an older managed runtime may never shadow a
+  // newer signed Desktop bundle after an app update.
+  eligible.sort((left, right) => {
+    if (!left.version && !right.version) return 0;
+    if (!left.version) return 1;
+    if (!right.version) return -1;
+    return compareSemVer(right.version, left.version);
+  });
+  return eligible[0]?.root ?? null;
 }
 
 function resolveCoreRuntimeRoot(explicitRoot, requiredMarkers = [], options = {}) {

package/engine/agentlas-i18n.cjs

@@ -21,7 +21,7 @@ const STRINGS = {
     "banner.location": "%s · / commands · Shift-Tab permissions",
     "banner.menu.help": "Command reference",
     "banner.menu.sessions": "Running sessions",
-    "banner.menu.chats": "Past conversations",
+    "banner.menu.chats": "Project runs",
     "banner.menu.quit": "Quit",
     "status.title": "Session",
     "status.runtime": "runtime",
@@ -93,10 +93,10 @@ const STRINGS = {
     "permFullConfirm": "FULL ACCESS ON FOR THIS SESSION — approvals and sandboxing are bypassed",
     "sideUsage": "usage: /side <question>",
     "sideNeedsSubject": "pick an agent first, or type the task normally so Agentlas can auto-route it",
-    "sideStart": "side question; this answer will not be saved to chat context",
-    "sideDone": "side question complete; main chat context unchanged",
-    "sideCancelled": "side question cancelled; main chat context unchanged",
-    "sideFailed": "side question failed; main chat context unchanged",
+    "sideStart": "side question; this answer will not be saved to the project task",
+    "sideDone": "side question complete; project task context unchanged",
+    "sideCancelled": "side question cancelled; project task context unchanged",
+    "sideFailed": "side question ended; project task context unchanged",
     "noAgent": "no agent: %s",
     "noCompany": "no company: %s",
     "agentUsage": "usage: /agent <name>",
@@ -139,7 +139,7 @@ const STRINGS = {
     "team.set": "%s → %s",
     "routedImage": "routed to %s for image support",
     "judge.source.llm": "judged by the connected model",
-    "judge.source.fallback": "deterministic fallback (no connected-model verdict)",
+    "judge.source.fallback": "unresolved (no model judgment)",
     "usageBar": "session usage",
     "config.title": "Engine auto engagement — explicit on/off (default: off)",
     "config.storm": "Stormbreaker auto-engage on direct-routed real work",
@@ -150,7 +150,7 @@ const STRINGS = {
     "config.autoEngage": "auto-engaging %s (opted in via /config — turn off with /config %s off)",
     "help.config": "explicit on/off for Stormbreaker/Workforce Ontology auto engagement",
     "guard.imageWarn": "this agent works with images, but %s can't generate them — try /runtime gemini (or codex), or /team to pin one",
-    "multimodal.title": "Multimodal fallback",
+    "multimodal.title": "Multimodal provider priority",
     "multimodal.set": "multimodal %s → %s",
     "multimodal.usage": "set with /multimodal set <image|video|audio> <provider-id>",
     // help rows
@@ -177,12 +177,12 @@ const STRINGS = {
     "help.memory": "show the memory being injected",
     "help.careerGraph": "show/add source refs for the Career Graph routing index",
     "help.ontology": "turn on/list/add project ontology sources with short natural commands",
-    "help.side": "ask a side question without saving it to chat context",
+    "help.side": "ask a side question without saving it to project task context",
     "help.status": "show model/runtime, agent, permission, and directory",
-    "help.multimodal": "show/set image, video, and audio fallback providers",
+    "help.multimodal": "show/set image, video, and audio provider priorities",
     "help.import": "import a local folder (agent or team)",
     "help.market": "browse/install marketplace agents",
-    "help.clear": "clear the chat and redraw",
+    "help.clear": "clear the terminal view and redraw",
     "help.storm": "run a force-robust Stormbreaker pipeline on a goal",
     "help.build": "build/repair/package an agent or team (Hephaestus)",
     "help.route": "preview which agent/pipeline would take a request",
@@ -272,7 +272,7 @@ const STRINGS = {
     "banner.location": "%s · / 명령 · Shift-Tab 권한",
     "banner.menu.help": "명령 보기",
     "banner.menu.sessions": "실행 중인 세션",
-    "banner.menu.chats": "지난 대화",
+    "banner.menu.chats": "프로젝트 실행",
     "banner.menu.quit": "종료",
     "status.title": "세션",
     "status.runtime": "런타임",
@@ -342,10 +342,10 @@ const STRINGS = {
     "permFullConfirm": "이 세션에서만 무제한 권한 켜짐 — 승인과 샌드박스를 우회함",
     "sideUsage": "사용법: /side <질문>",
     "sideNeedsSubject": "먼저 에이전트를 고르거나, 그냥 할 일을 입력해서 Agentlas가 자동 라우팅하게 하세요.",
-    "sideStart": "사이드 질문입니다. 이 답변은 메인 대화 맥락에 저장하지 않습니다.",
-    "sideDone": "사이드 질문 완료. 메인 대화 맥락은 그대로입니다.",
-    "sideCancelled": "사이드 질문이 중단되었습니다. 메인 대화 맥락은 그대로입니다.",
-    "sideFailed": "사이드 질문이 실패했습니다. 메인 대화 맥락은 그대로입니다.",
+    "sideStart": "사이드 질문입니다. 이 답변은 프로젝트 작업에 저장하지 않습니다.",
+    "sideDone": "사이드 질문 완료. 프로젝트 작업 맥락은 그대로입니다.",
+    "sideCancelled": "사이드 질문이 중단되었습니다. 프로젝트 작업 맥락은 그대로입니다.",
+    "sideFailed": "사이드 질문이 종료되었습니다. 프로젝트 작업 맥락은 그대로입니다.",
     "noAgent": "에이전트 없음: %s",
     "noCompany": "회사 없음: %s",
     "agentUsage": "사용법: /agent <이름>",
@@ -388,7 +388,7 @@ const STRINGS = {
     "team.set": "%s → %s",
     "routedImage": "이미지 지원을 위해 %s로 라우팅",
     "judge.source.llm": "판정: 연결 모델",
-    "judge.source.fallback": "판정: 결정적 폴백(연결 모델 없음)",
+    "judge.source.fallback": "판정 미확정(모델 판단 없음)",
     "usageBar": "세션 사용량",
     "config.title": "엔진 자동 개입 — 명시적 on/off (기본: off)",
     "config.storm": "직답 라우팅된 실작업에 Stormbreaker 자동 개입",
@@ -399,7 +399,7 @@ const STRINGS = {
     "config.autoEngage": "%s 자동 개입 (/config로 켜져 있음 — 해제: /config %s off)",
     "help.config": "Stormbreaker/Workforce Ontology 자동 개입 명시적 on/off",
     "guard.imageWarn": "이 에이전트는 이미지 작업인데 %s는 이미지 생성 불가 — /runtime gemini(또는 codex), 또는 /team으로 고정",
-    "multimodal.title": "멀티모달 대체 제공자",
+    "multimodal.title": "멀티모달 제공자 우선순위",
     "multimodal.set": "멀티모달 %s → %s",
     "multimodal.usage": "/multimodal set <image|video|audio> <provider-id> 로 변경",
     "help.talk": "현재 에이전트/회사와 대화 — 스트리밍 + 툴",
@@ -425,12 +425,12 @@ const STRINGS = {
     "help.memory": "주입되는 메모리 보기",
     "help.careerGraph": "경력 그래프 라우팅 색인의 원본 경로 보기/등록",
     "help.ontology": "짧은 자연어로 프로젝트 온톨로지 켜기/목록/자료 등록",
-    "help.side": "메인 대화 맥락에 저장하지 않는 사이드 질문",
+    "help.side": "프로젝트 작업 맥락에 저장하지 않는 사이드 질문",
     "help.status": "모델/런타임, 에이전트, 권한, 작업 폴더 보기",
-    "help.multimodal": "이미지·영상·음성 대체 제공자 보기/변경",
+    "help.multimodal": "이미지·영상·음성 제공자 우선순위 보기/변경",
     "help.import": "로컬 폴더(에이전트/팀) 임포트",
     "help.market": "마켓플레이스 에이전트 보기/설치",
-    "help.clear": "대화 비우고 다시 그리기",
+    "help.clear": "터미널 화면 비우고 다시 그리기",
     "help.storm": "목표를 Stormbreaker 견고 파이프라인으로 실행",
     "help.build": "에이전트/팀 빌드·수리·패키징 (Hephaestus)",
     "help.route": "요청이 어떤 에이전트로 라우팅되는지 미리보기",
@@ -443,7 +443,7 @@ const STRINGS = {
     "help.connect": "Telegram/플랫폼을 에이전트 팀에 연결",
     "help.swarm": "목표에 창발형 에이전트 스웜 전개",
     "help.doctor": "런타임/데이터 점검",
-    "help.compact": "오래된 대화 맥락을 줄이고 최근 맥락 유지",
+    "help.compact": "오래된 작업 맥락을 줄이고 최근 맥락 유지",
     "help.keybindings": "터미널 단축키 보기",
     "help.exit": "종료 (Ctrl-C: 턴 중단 / 유휴 시 종료)",
     "help.talkKey": "(메시지 입력)",

package/engine/agentlas-native-host.cjs

@@ -964,7 +964,6 @@ function runNativeTurn(req) {
         return;
       }
       ui.stopSpinner();
-      ui.error(uiText(ui, "runtime.failed", kind, err.message));
       finish({ text: "", session: st.session, error: err.message });
     });
     child.on("close", (code) => {
@@ -976,39 +975,13 @@ function runNativeTurn(req) {
       ui.streamEnd();
       ui.stopSpinner();
       const text = (st.finalText || st.text || "").trim();
-      const errTail = stripAnsi(stderrBuf).replace(/\s+/g, " ").trim(); // ANSI 제거 + 한 줄로
-      if (st.error && !st.errorShown) {
-        // claude `result` is_error 등 — 이전에 표시되지 않은 에러를 노출
-        ui.error(String(st.error));
-      } else if (code !== 0) {
-        // Runtime Doctor — 아는 시스템 원인(미인증 OAuth MCP 플러그인 등)이면 즉시 수리하고
-        // 1회 자동 재시도한다(2026-07-08 notion@openai-curated가 codex 전멸시킨 사고).
-        if (!req._doctorRetried) {
-          try {
-            const { runRuntimeDoctor } = require("./agentlas-doctor.cjs");
-            const report = runRuntimeDoctor(`${kind} exited with code ${code}\n${stripAnsi(stderrBuf)}`);
-            if (report.repaired) {
-              ui.warn(uiText(ui, "runtime.doctor", report.summary));
-              for (const act of report.actions) ui.warn(`   ${act.title} — ${act.detail}`);
-              ui.warn(uiText(ui, "runtime.doctorRetry"));
-              settled = true;
-              cleanup();
-              resolve(runNativeTurn({ ...req, _doctorRetried: true }));
-              return;
-            }
-          } catch {
-            /* 닥터 실패는 원래 에러 표출을 막지 않는다 */
-          }
-        }
-        ui.error(uiText(ui, "runtime.exited", kind, String(code)) + (errTail ? `\n  ${errTail.slice(-400)}` : ""));
-      } else if (!text && !st.error) {
-        // 정상 종료인데 출력이 비어 있음(거부/차단 등) — 무음 실패 방지
-        ui.warn(uiText(ui, "runtime.noOutput", kind) + (errTail ? ` (${errTail.slice(-200)})` : ""));
-      }
       if (st.usage) ui.cost(st.usage);
       const exitError = code === 0
         ? st.error
-        : st.error || `${kind} exited with code ${code == null ? "unknown" : code}`;
+        : st.error || [
+            `${kind} exited with code ${code == null ? "unknown" : code}`,
+            stripAnsi(stderrBuf).slice(-4000),
+          ].filter(Boolean).join("\n");
       finish({ text, session: st.session, usage: st.usage, error: exitError });
     });
 

package/engine/agentlas-ui.cjs

@@ -594,7 +594,19 @@ class Ui {
   }
   error(msg) {
     this.stopSpinner();
-    this._message("✗ ", this.c.paw, this.c.text, msg);
+    // Last-resort presentation boundary for legacy/direct commands. Raw
+    // provider text, stack messages, paths and codes must never become UI.
+    // REPL/session paths route the private evidence to the controller before
+    // reaching this boundary; direct commands get a neutral recovery state.
+    void msg;
+    this._message(
+      "◆ ",
+      this.c.amber,
+      this.c.text,
+      this.lang === "ko"
+        ? "One이 상태를 확인하고 복구하고 있습니다."
+        : "One is checking and recovering this operation.",
+    );
   }
 
   // 최종 텍스트(비스트리밍 경로)에 가벼운 마크다운 강조 적용 후 출력.

package/engine/agentlas-workload-routing.cjs

@@ -308,7 +308,7 @@ function resolveAllocation(options = {}) {
       model: modelPin,
       effort: pinnedEffort,
       provider,
-      source: modelPin || effortPinPresent ? "user-pin" : "fallback",
+      source: modelPin ? "user-pin" : "unresolved",
       fallbackReason: pinReasons.join(","),
       aiReason: null,
     };
@@ -355,24 +355,14 @@ function resolveAllocation(options = {}) {
 
   let model = selected && selected.id;
   let effectiveModelEntry = selected;
-  let source = decision.exactModelId ? "parent-ai-exact" : "fallback";
+  let source = decision.exactModelId ? "parent-ai-exact" : "unresolved";
   if (modelPin) {
     model = modelPin;
     effectiveModelEntry = available.find((item) => item.id === modelPin) || null;
     source = "user-pin";
     reasons.push("explicit_model_pin");
   } else if (!model) {
-    source = "fallback";
-    const activeId = cleanText(runtime && runtime.model, 160) || null;
-    const active = activeId ? available.find((item) => item.id === activeId) || null : null;
-    const issue = activeId ? candidateIssue(active, "active_model") : "active_model_unavailable";
-    if (issue) {
-      if (!reasons.includes(issue)) reasons.push(issue);
-    } else {
-      model = active.id;
-      effectiveModelEntry = active;
-      reasons.push("compliant_active_model_fallback");
-    }
+    source = "unresolved";
   }
 
   let effort;
@@ -406,21 +396,20 @@ function resolveAllocation(options = {}) {
 function resolveAllocationAcrossRuntimes(options = {}) {
   const runtimes = Array.isArray(options.runtimes) ? options.runtimes : [];
   const decision = normalizeAllocation(options.decision);
-  const fallbackRuntime = options.fallbackRuntime || options.runtime || runtimes[0] || null;
-  const fallbackId = cleanText(fallbackRuntime && (fallbackRuntime.runtimeId || fallbackRuntime.id), 255) || null;
+  const explicitRuntime = options.runtime || null;
   const requestedId = decision && decision.runtimeId;
   const chosen = requestedId
     ? runtimes.find((runtime, index) => (cleanText(runtime && (runtime.runtimeId || runtime.id), 255) || `runtime-${index + 1}`) === requestedId) || null
-    : fallbackRuntime;
-  const runtime = chosen || fallbackRuntime;
+    : explicitRuntime;
+  const runtime = chosen || null;
   const resolution = resolveAllocation({ ...options, runtime, decision, availableModels: runtime && runtime.availableModels });
   const requestedExact = Boolean(decision && decision.runtimeId && decision.exactModelId);
-  const runtimeId = cleanText(runtime && (runtime.runtimeId || runtime.id), 255) || fallbackId;
+  const runtimeId = cleanText(runtime && (runtime.runtimeId || runtime.id), 255) || null;
   if (requestedExact && chosen && resolution.model === decision.exactModelId) {
     resolution.source = "parent-selected-live-runtime-model";
   } else if (requestedExact) {
     resolution.fallbackReason = [resolution.fallbackReason, chosen ? "parent_model_not_in_live_inventory" : "parent_runtime_not_in_live_inventory"].filter(Boolean).join(",");
-    if (resolution.source !== "user-pin") resolution.source = "fallback";
+    if (resolution.source !== "user-pin") resolution.source = "unresolved";
   }
   return { ...resolution, runtime, runtimeId, requestedRuntimeId: requestedId || null };
 }
@@ -478,9 +467,7 @@ function createDecisionReceipt({ taskId, stage, decision, resolution, role, usag
   );
   const status = source === "user-pin" && hasResolvedCurrent
     ? "user-pin"
-    : source === "fallback" && hasResolvedCurrent
-      ? "fallback-current"
-      : normalized && resolution && resolution.ok
+    : normalized && resolution && resolution.ok
         ? "resolved"
         : "unresolved";
   const riskCodes = new Set(normalized ? normalized.reasonCodes : []);
@@ -509,7 +496,7 @@ function createDecisionReceipt({ taskId, stage, decision, resolution, role, usag
     },
     reasonCodes,
     inputFeatureHash: normalized && normalized.inputFeatureHash ? normalized.inputFeatureHash : featureHash,
-    selectorVersion: normalized ? normalized.selectorVersion : "deterministic-host-fallback",
+    selectorVersion: normalized ? normalized.selectorVersion : "unresolved-no-model-judgment",
     independentVerificationRequired:
       riskCodes.has("high-risk") || riskCodes.has("critical-risk") || riskCodes.has("independent-verification"),
     usage: normalizeObservedUsage(usage),
@@ -548,7 +535,7 @@ function plannerSystemPrompt({ language = "English", maxTasks = 12, mode = "swar
   return [
     `You are the higher-level workload allocator for an Agentlas ${mode}.`,
     "Judge each child task using the full goal and planned dependency graph. Do not use a keyword lookup or fixed role-to-model table.",
-    "LIVE_RUNTIME_INVENTORY below is authoritative. For every child and synthesis, choose an exact runtimeId and exactModelId only from it. Do not infer, rename, or invent a model from a tier. If an exact choice cannot be justified, select the current-runtime fallback from the inventory.",
+    "LIVE_RUNTIME_INVENTORY below is authoritative. For every child and synthesis, choose an exact runtimeId and exactModelId only from it. Do not infer, rename, invent, or substitute a model. If an exact choice cannot be justified, return an unresolved decision.",
     `LIVE_RUNTIME_INVENTORY=${JSON.stringify(liveRuntimeInventory)}`,
     "Choose effort none|minimal|low|medium|high|xhigh|max. Spend frontier/high effort only when the task's complexity, risk, context, or synthesis burden justifies it.",
     `Return strict JSON only with at most ${Math.max(1, Math.min(24, maxTasks))} tasks:`,