@su-record/vibe 2.12.2 → 2.12.4

package/hooks/scripts/llm-orchestrate.js

@@ -42,7 +42,9 @@ const mode = process.argv[3] || 'orchestrate';
 // WHY 3 retries: Enough to ride out brief 503/overload blips (typically 1-2
 // consecutive), but not so many that a genuinely down provider delays the
 // fallback chain for minutes.
-const MAX_RETRIES = 3;
+// VIBE_LLM_MAX_RETRIES override: UserPromptSubmit(hook) 모드에서는 1(재시도 없음)로
+// 낮춰 부모 dispatcher timeout 안에 단일 시도로 끝낸다. (B-2: hard-kill/무음실패 방지)
+const MAX_RETRIES = Number(process.env.VIBE_LLM_MAX_RETRIES) || 3;
 // WHY 2000ms initial delay: LLM rate-limit windows are typically 1-5s;
 // starting at 2s with exponential backoff (2s, 4s, 8s) covers most reset intervals.
 const INITIAL_DELAY_MS = 2000;
@@ -203,7 +205,9 @@ function parseAnalyzeImageArgs(args) {
 // CLI Provider Functions
 // ============================================
 
-const CLI_TIMEOUT_MS = 180000;
+// VIBE_LLM_PRIMARY_TIMEOUT_MS override: hook 모드에서는 부모 dispatcher timeout 보다
+// 짧게 잡아(예: 45s) 자식이 스스로 정리하고 의미있는 메시지를 반환하게 한다.
+const CLI_TIMEOUT_MS = Number(process.env.VIBE_LLM_PRIMARY_TIMEOUT_MS) || 180000;
 const CLI_FALLBACK_TIMEOUT_MS = 30000;
 const IS_WINDOWS = os.platform() === 'win32';
 
@@ -599,6 +603,13 @@ async function main() {
     providerChain = claudeSecondary ? ['antigravity', 'claude'] : ['antigravity', 'gpt'];
   }
 
+  // hook(UserPromptSubmit) 모드: 사용자가 `gpt`/`agy` 접두사로 명시적으로 부른 단발
+  // 보조 호출이라 cross-provider fallback 이 불필요하다. primary 1개로 단축해
+  // 부모 dispatcher timeout 안에 단일 시도로 끝낸다. (B-2)
+  if (process.env.VIBE_LLM_HOOK_MODE) {
+    providerChain = [providerChain[0]];
+  }
+
   const vibeConfig = readVibeConfig();
 
   for (let i = 0; i < providerChain.length; i++) {

package/hooks/scripts/prompt-dispatcher.js

@@ -82,39 +82,47 @@ const DISPATCH_RULES = [
     label: 'e2e-echo',
   },
 
-  // 외부 LLM 호출 (GPT/Antigravity) - 패턴 매칭 필수
+  // 외부 LLM 호출 (GPT/Antigravity) — 명시적 provider 접두사 필수.
+  //
+  // 과거에는 "추론해", "코드 리뷰", "디버깅해" 같은 자연어 패턴이 prompt
+  // 어디서든 매칭되어, 평범한 한국어/영어 요청에도 외부 LLM이 동기로 spawn되고
+  // 그 응답이 컨텍스트에 주입되었다(컨텍스트 오염 + 최대 30s 블로킹).
+  // 이제는 prompt가 `gpt`/`agy`/`antigravity` 로 **시작**할 때만 발동한다.
+  // 즉 사용자가 외부 LLM을 콕 집어 부를 때만 동작하고, 일상 요청엔 걸리지 않는다.
+  // (참고: `/vibe.reason` 스킬이 일반 추론을 담당하므로 자연어 자동호출은 불필요)
+  // `s` 플래그(dotAll)로 여러 줄 prompt 의 역할 키워드도 매칭한다.
   {
-    pattern: /아키텍처.*(검토|리뷰|분석)|architecture.*(review|analyz)|설계.*검토|구조.*분석.*해/i,
+    pattern: /^\s*gpt\b.*(아키텍처|architecture|설계|구조)/is,
     script: 'llm-orchestrate.js',
     args: ['gpt', 'orchestrate', 'You are a software architect. Analyze and review the architecture.'],
     label: 'gpt-architecture',
   },
   {
-    pattern: /(UI|UX).*(리뷰|검토|피드백|개선)|사용자.*경험.*검토|디자인.*리뷰|design.*feedback/i,
+    pattern: /^\s*(agy|antigravity)\b.*(ui|ux|디자인|design|사용자.*경험)/is,
     script: 'llm-orchestrate.js',
     args: ['antigravity', 'orchestrate', 'You are a UI/UX expert. Analyze and provide feedback.'],
     label: 'antigravity-uiux',
   },
   {
-    pattern: /디버깅.*해|버그.*찾아|find.*bug|debug.*this.*code/i,
+    pattern: /^\s*gpt\b.*(디버깅|debug|버그|bug)/is,
     script: 'llm-orchestrate.js',
     args: ['gpt', 'orchestrate', 'You are a debugging expert. Find bugs and suggest fixes.'],
     label: 'gpt-debug',
   },
   {
-    pattern: /코드.*정적.*분석|코드.*분석.*해줘|analyze.*code.*quality/i,
+    pattern: /^\s*(agy|antigravity)\b.*(분석|analyz|코드.*품질|code.*quality)/is,
     script: 'llm-orchestrate.js',
     args: ['antigravity', 'orchestrate', 'You are a code analysis expert. Review and analyze the code.'],
     label: 'antigravity-analysis',
   },
   {
-    pattern: /코드.*리뷰|code.*review|PR.*리뷰|리뷰.*해줘.*코드/i,
+    pattern: /^\s*gpt\b.*(리뷰|review)/is,
     script: 'llm-orchestrate.js',
     args: ['gpt', 'orchestrate', 'You are a code review expert. Review the code for best practices, security, and performance.'],
     label: 'gpt-codereview',
   },
   {
-    pattern: /추론.*해|reasoning|복잡.*분석|deep.*analysis/i,
+    pattern: /^\s*gpt\b.*(추론|reasoning|복잡.*분석|deep.*analysis)/is,
     script: 'llm-orchestrate.js',
     args: ['gpt', 'orchestrate', 'You are a reasoning expert. Analyze the problem deeply and provide detailed reasoning.'],
     label: 'gpt-reasoning',
@@ -155,14 +163,29 @@ for (const rule of DISPATCH_RULES) {
   const scriptPath = path.join(__dirname, rule.script);
   const args = rule.args || [];
 
+  // 외부 LLM 호출(llm-orchestrate)은 부모/자식 timeout 을 정합시킨다 (B-2):
+  // 자식은 hook 모드(primary 45s, fallback 없음, retry 없음)로 단일 시도 후 스스로
+  // 정리하고, 부모는 그보다 약간 긴 50s 로 감싸 hard-kill 을 피한다.
+  // 경량 스크립트(keyword-detector 등)는 기존 30s 유지.
+  const isLlm = rule.script === 'llm-orchestrate.js';
+  const execTimeout = isLlm ? 50000 : 30000;
+  const childEnv = isLlm
+    ? { ...process.env, VIBE_LLM_HOOK_MODE: '1', VIBE_LLM_PRIMARY_TIMEOUT_MS: '45000', VIBE_LLM_MAX_RETRIES: '1' }
+    : { ...process.env };
+
   execPromises.push(
     new Promise((resolve) => {
       execFile('node', [scriptPath, ...args], {
-        timeout: 30000,
-        env: { ...process.env },
+        timeout: execTimeout,
+        env: childEnv,
       }, (error, stdout, stderr) => {
         if (stdout?.trim()) {
           process.stdout.write(stdout);
+        } else if (error) {
+          // 무음실패 방지 — 외부 LLM 호출이 timeout 또는 에러로 결과 없이 죽을 때,
+          // 사용자가 원인을 알 수 있도록 한 줄만 노출한다.
+          const reason = error.killed ? `timed out (${execTimeout / 1000}s)` : (error.message || 'failed');
+          process.stdout.write(`[${rule.label}] external LLM call ${reason} — no result injected.\n`);
         }
         resolve();
       });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@su-record/vibe",
-  "version": "2.12.2",
+  "version": "2.12.4",
   "description": "AI Coding Framework for Claude Code — 56 agents, 45 skills, multi-LLM orchestration",
   "type": "module",
   "main": "dist/cli/index.js",

package/skills/design-audit/SKILL.md

@@ -1,5 +1,6 @@
 ---
 name: design-audit
+user-invocable: true
 invocation: [command, auto]
 tier: standard
 description: "Design technical quality audit — a11y, performance, responsive, theming, AI slop detection with 5-dimension scoring. Use when design-audit, ui-audit, a11y-check, design-check."

package/skills/design-normalize/SKILL.md

@@ -1,5 +1,6 @@
 ---
 name: design-normalize
+user-invocable: true
 invocation: [command, auto]
 tier: standard
 description: "Normalize hardcoded values to design system tokens — colors, typography, spacing, shadows aligned to MASTER.md. Use when design-normalize, design-system, token-align."

package/skills/design-teach/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: design-teach
 user-invocable: false
-invocation: [command, auto]
+invocation: [auto]
 tier: standard
 description: "Gather and store project design context — target audience, brand personality, aesthetic direction, constraints. Used by other design-* skills. Use when design-teach, design-setup, design-context."
 triggers: [design-teach, design-setup, design-context]

package/skills/test/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: test
-description: vibe 자가검진 본체 — 대상 harness(~/.claude/~/.codex)의 모든 command/skill/hook/agent 프로빙 → pass/fail 리포트 → ~/.vibe/test-reports/.
-when_to_use: /vibe.test 진입점에서 체인 호출. 직접 호출 금지.
+description: vibe 자가검진 본체 — 대상 harness(~/.claude/~/.codex)의 모든 entry skill/skill/hook/agent 프로빙 → pass/fail 리포트 → ~/.vibe/test-reports/.
+when_to_use: vibe.test 진입점에서 체인 호출. 직접 호출 금지.
 user-invocable: false
 tier: core
 ---
@@ -12,7 +12,7 @@ Probe every shipped vibe surface in one install dir and emit a pass/fail report.
 
 ## Why this exists
 
-When vibe ships new commands, skills, hooks, or agents, one side (CC or Codex) can end up out of sync with the other, frontmatter can drift, and hook tests can silently break. `/vibe.test` is the single mechanical check: does every surface in the target install actually load and pass its own tests?
+When vibe ships new entry skills, skills, hooks, or agents, one side (CC or Codex) can end up out of sync with the other, frontmatter can drift, and hook tests can silently break. `vibe.test` is the single mechanical check: does every surface in the target install actually load and pass its own tests?
 
 ## Target harness
 
@@ -37,8 +37,8 @@ All probes are **structural or test-based** — no interactive command is ever a
 
 | Category | Source | Check |
 |---|---|---|
-| commands | `<install>/commands/*.md` | file readable · frontmatter parses · `description` present · body references a skill (`Load skill \`...\``) if it delegates |
-| skills | `<install>/skills/*/SKILL.md` | frontmatter parses · required fields (`name`, `description`, `invocation`) · `invocation` ⊆ `{command, auto, chain}` and non-empty · if `invocation` includes `auto` then `triggers` non-empty · if includes `command` then matching `commands/{vibe.X}.md` or `Load skill <name>` reference exists · if includes `chain` then another skill lists it in `chain-next` · body non-empty |
+| entry skills | `<install>/skills/vibe*/SKILL.md` | file readable · frontmatter parses · `name`, `description`, `user-invocable: true` present |
+| skills | `<install>/skills/*/SKILL.md` | frontmatter parses · required fields (`name`, `description`) · body non-empty |
 | hooks | repo `hooks/scripts/*.js` | for each script with a matching `__tests__/<name>.test.js`, run `npx vitest run <test> --reporter=json` and parse pass/fail counts |
 | agents | `<install>/agents/*.md` | file readable · frontmatter parses · required fields (`name`, `description`) |
 
@@ -62,7 +62,7 @@ Written to `~/.vibe/test-reports/<YYYYMMDD-HHmm>-<harness>.{json,md}`. Exact sch
     "failed": 2
   },
   "probes": {
-    "commands": [
+    "entrySkills": [
       { "name": "vibe.spec", "status": "pass" },
       { "name": "vibe.test", "status": "pass" }
     ],
@@ -102,7 +102,7 @@ Written to `~/.vibe/test-reports/<YYYYMMDD-HHmm>-<harness>.{json,md}`. Exact sch
 
 | Category | Pass | Fail |
 |---|---:|---:|
-| commands | 15 | 0 |
+| entry skills | 15 | 0 |
 | skills   | 17 | 1 |
 | hooks    |  6 | 0 |
 | agents   |  3 | 1 |
@@ -136,5 +136,5 @@ If `failed` is empty, replace the Failures section with `_All probes passed._`.
 - [ ] JSON report matches the template above exactly (fields, types, naming)
 - [ ] Markdown summary printed to console after the run
 - [ ] Reports land in `~/.vibe/test-reports/`, never in project-local `.vibe/`
-- [ ] `failed.length > 0` → auto-invokes `/vibe.regress register --from-test`
-- [ ] `invocation` field invariants enforced per skill (`command` / `auto` / `chain` validity)
+- [ ] `failed.length > 0` → auto-invokes `vibe.regress register --from-test`
+- [ ] Entry skills are verified as user-invocable skill surfaces, not deprecated command files

package/skills/ui-ux-pro-max/SKILL.md

@@ -1,5 +1,6 @@
 ---
 name: ui-ux-pro-max
+user-invocable: true
 invocation: [command]
 tier: standard
 description: "UI/UX design intelligence with bold aesthetic direction. Searchable database: 50+ styles, 97 color palettes, 57 font pairings, 99 UX guidelines, 25 chart types across 9 stacks (React, Vue, Svelte, Angular, Astro, Next.js, Flutter, Swift, Kotlin). Use when building any UI component, choosing color schemes, selecting typography, applying design patterns, or reviewing UX compliance. Must use this skill when user builds frontend — pages, components, dashboards, or landing pages — to ensure design quality beyond generic AI output."

package/skills/vibe.analyze/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: vibe.analyze
 description: Analyze any target — code, document, website, or Figma design
-argument-hint: "feature-name" or file.pdf or https://... or --code or --deps or --arch
+argument-hint: '"feature-name" or file.pdf or https://... or --code or --deps or --arch'
 user-invocable: true
 ---
 

package/skills/vibe.event/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: vibe.event
 description: Community event automation — D-Day based proactive task execution
-argument-hint: "MDC 12차 3/30" or "webinar AI에이전트 4/15" or "DWK 5차 5/20"
+argument-hint: '"MDC 12차 3/30" or "webinar AI에이전트 4/15" or "DWK 5차 5/20"'
 user-invocable: true
 ---
 

package/skills/vibe.run/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: vibe.run
 description: Execute implementation from SPEC
-argument-hint: "feature name" or --phase N
+argument-hint: '"feature name" or --phase N'
 user-invocable: true
 ---
 
@@ -606,7 +606,7 @@ At automation levels L3 and L4, checkpoints that do not require confirmation are
 
 - `core/development-philosophy.md` - Surgical precision, modify only requested scope
 - `core/quick-start.md` - Korean, DRY, SRP, YAGNI
-- `standards/complexity-metrics.md` - Functions ≤20 lines, nesting ≤3 levels
+- `standards/complexity-metrics.md` - Functions ≤50 lines, nesting ≤3 levels
 - `quality/checklist.md` - Code quality checklist
 
 **Language guide:** `~/.claude/vibe/languages/{stack}.md` (global reference)
@@ -1617,7 +1617,7 @@ After all scenarios complete + Antigravity review, **quality report is auto-gene
 │  │ Build             │ ✅     │ npm run build succeeded     │    │
 │  │ Tests             │ ✅     │ 12/12 passed                │    │
 │  │ Type check        │ ✅     │ 0 errors                    │    │
-│  │ Complexity        │ ✅     │ All functions ≤30 lines     │    │
+│  │ Complexity        │ ✅     │ All functions ≤50 lines     │    │
 │  │ Security          │ ✅     │ 0 vulnerabilities           │    │
 │  │ Race review       │ ✅     │ 3 improvements applied      │    │
 │  └─────────────────────────────────────────────────────────┘    │
@@ -1857,7 +1857,7 @@ Before marking any scenario as complete, ALL items must pass:
 | **Functionality** | All Given/When/Then conditions verified | 20% |
 | **Functionality** | Edge cases handled per scenario | 10% |
 | **Code Quality** | No `any` types in TypeScript | 10% |
-| **Code Quality** | Functions ≤30 lines, nesting ≤3 levels | 10% |
+| **Code Quality** | Functions ≤50 lines, nesting ≤3 levels | 10% |
 | **Code Quality** | No hardcoded values (use constants) | 5% |
 | **Security** | Input validation implemented | 10% |
 | **Security** | Authentication/authorization checked | 5% |

package/skills/vibe.test/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: vibe.test
-description: Self-test vibe — probe every command/skill/hook/agent in the target harness install dir and write a pass/fail report
+description: Self-test vibe — probe every entry skill/skill/hook/agent in the target harness install dir and write a pass/fail report
 argument-hint: "[cc|codex]  (empty = current harness)"
 user-invocable: true
 ---

package/skills/vibe.verify/SKILL.md

@@ -257,7 +257,7 @@ Record the value as `{step_count}` and include it in the report below.
 │  │ Build             │ ✅    │ Success                     │    │
 │  │ Tests             │ ✅    │ 12/12 passed                │    │
 │  │ Type Check        │ ✅    │ 0 errors                    │    │
-│  │ Complexity        │ ✅    │ All functions ≤30 lines     │    │
+│  │ Complexity        │ ✅    │ All functions ≤50 lines     │    │
 │  │ Code Coverage     │ ⚠️    │ 78% (target: 80%)           │    │
 │  └─────────────────────────────────────────────────────────┘    │
 │                                                                 │

package/vibe/rules/standards/complexity-metrics.md

@@ -283,7 +283,7 @@ npm install eslint-plugin-complexity
   "rules": {
     "complexity": ["error", 10],
     "max-depth": ["error", 3],
-    "max-lines-per-function": ["error", 20]
+    "max-lines-per-function": ["error", 50]
   }
 }
 ```
@@ -307,7 +307,7 @@ radon mi app/
 |--------|--------|-------------|
 | Cyclomatic Complexity | ≤ 10 | Independent execution paths |
 | Cognitive Complexity | ≤ 15 | Easy to understand |
-| Function Length | ≤ 20 lines | Short, focused functions |
+| Function Length | ≤ 50 lines | Short, focused functions |
 | Nesting Depth | ≤ 3 levels | Flat structure |
 | Parameters | ≤ 5 | Function parameter limit |
 | Dependencies | ≤ 7 | Module dependency limit |