@su-record/vibe 3.2.5 → 3.2.7

package/skills/vibe.scaffold/SKILL.md

@@ -1,12 +1,19 @@
 ---
 name: vibe.scaffold
-description: Generate or audit project folder structure optimized for AI-assisted development
+description: 새 프로젝트 구조를 만들거나 기존 폴더 구조가 AI-assisted development에 적합한지 감사할 때 사용한다.
 argument-hint: --check (audit existing) or project-type
 user-invocable: true
 ---
 
 # /vibe.scaffold
 
+## 완료 기준
+
+- [ ] 요청된 디렉터리와 필수 설정 파일이 존재한다.
+- [ ] 생성한 package 또는 module의 build 명령이 통과한다.
+- [ ] import·module 해석 검사가 통과한다.
+- [ ] audit mode에서는 발견 항목마다 실제 경로 근거가 있다.
+
 Design and generate a project structure where AI works effectively on its own.
 
 > "A well-designed project structure trains AI to follow it naturally."
@@ -103,18 +110,7 @@ lib/
 
 ### 5. Clean Architecture Layers (Optional)
 
-```
-src/
-├── domain/           # Business rules (pure logic, no external deps)
-├── application/      # Use cases (domain composition)
-├── infrastructure/   # External integrations (DB, API, files)
-└── presentation/     # UI or API endpoints
-```
-
-Layer rules:
-- Dependency direction: presentation → application → domain (reverse forbidden)
-- domain must not import external packages
-- infrastructure implements domain interfaces
+When clean architecture is explicitly selected, read `references/clean-architecture.md`. Default stack layouts and `--check` audits do not load it.
 
 ### 6. Generate Supporting Files
 

package/skills/vibe.scaffold/references/clean-architecture.md

@@ -0,0 +1,17 @@
+# Optional Clean Architecture Layout
+
+Load only when clean architecture is explicitly selected.
+
+```
+src/
+├── domain/           # Business rules (pure logic, no external deps)
+├── application/      # Use cases (domain composition)
+├── infrastructure/   # External integrations (DB, API, files)
+└── presentation/     # UI or API endpoints
+```
+
+Layer rules:
+
+- Dependency direction: presentation → application → domain (reverse forbidden)
+- domain must not import external packages
+- infrastructure implements domain interfaces

package/skills/vibe.spec/SKILL.md

@@ -1,12 +1,19 @@
 ---
 name: vibe.spec
-description: SPEC 진입점 — 자연어 요구사항(+첨부)을 받아 단일 패스 SPEC 작성 → 1회 승인 → /vibe.run 핸드오프
+description: 자연어 요구사항이나 첨부 자료로 승인 가능한 SPEC을 작성해야 할 때 — 단일 패스로 작성하고 1회 승인 후 vibe.run으로 넘긴다.
 argument-hint: "(선택) feature name, requirement, file path, or idea"
 user-invocable: true
 ---
 
 # /vibe.spec
 
+## 완료 기준
+
+- [ ] `.vibe/specs/<feature>.md`가 존재한다.
+- [ ] 모든 REQ에 결정론적 Done Criteria와 Evidence Required가 있다.
+- [ ] 제약, 범위 제외, stakes가 명시되어 있다.
+- [ ] confirm mode에서는 사용자 승인 상태가 기록되어 있다.
+
 **SPEC 단계의 얇은 진입점.** 요구사항을 받아 `spec` 스킬(단일 패스)을 실행하고, SPEC 승인(유일한 의무 게이트) 후 `/vibe.run` 으로 넘긴다. interview → plan → spec → review 다단계 파이프라인은 폐지되었다 — 명확화 질문은 spec 패스 안에서 인라인으로, 리뷰는 셀프 리뷰 1회로 흡수됐다.
 
 ## Usage
@@ -50,10 +57,7 @@ SPEC 승인 (1회 — automationLevel: autonomous 면 생략)
 
 ### 레거시 아티팩트 호환 (구버전 .vibe/interviews/, .vibe/plans/)
 
-구버전 vibe 가 남긴 `.vibe/interviews/{feature}.md`, `.vibe/plans/{feature}.md` 가 존재할 수 있다:
-
-- **존재하면**: spec 패스의 **입력 컨텍스트**로 읽어 활용한다 (이미 답한 질문을 다시 묻지 않게).
-- **절대 요구하거나 재생성하지 않는다**: 이 파일들의 부재는 결함이 아니라 정상 상태다. interview/plan 스킬은 더 이상 존재하지 않는다.
+Legacy `.vibe/interviews/` or `.vibe/plans/` artifacts가 실제로 감지된 경우에만 `references/legacy-artifacts.md`를 읽는다. 신규 프로젝트와 현재 형식만 있는 호출은 읽지 않는다.
 
 ### `.last-feature` 포인터
 
@@ -109,11 +113,11 @@ ARGUMENTS: $ARGUMENTS
 ### 2. Project context
 
 - `.vibe/config.json` 읽기 — `references.languages[]` 의 스택 가이드, `stacks` 확인.
-- 기존 코드 파악이 필요하면 네이티브 Explore 서브에이전트에 위임 (main session 에서 프로젝트 파일을 훑지 않는다):
+- 기존 코드 파악이 필요하면 하네스의 네이티브 협업 기능에 독립 탐색 worker로 위임한다. Claude Code는 Task/Agent, Codex는 native collaboration에 매핑하며 기본 모델을 상속한다 (coordinator session에서 프로젝트 전체를 훑지 않는다):
 
 ```text
-Task(subagent_type="Explore",
-  prompt="Find existing implementations related to [FEATURE]. Return: tech stack, relevant files, patterns. Under 200 tokens.")
+Worker: Find existing implementations related to [FEATURE]. Return the tech stack,
+relevant files, and patterns in under 200 tokens.
 ```
 
 ### 3. Clarify — 진짜 모호할 때만

package/skills/vibe.spec/references/legacy-artifacts.md

@@ -0,0 +1,6 @@
+# Legacy SPEC Input Artifacts
+
+Load only when `.vibe/interviews/{feature}.md` or `.vibe/plans/{feature}.md` exists.
+
+- Read existing legacy artifacts as SPEC input context so answered questions are not repeated.
+- Never require or regenerate them. Their absence is normal; interview/plan skills no longer exist.

package/skills/vibe.test/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: vibe.test
-description: Self-test vibe — probe every entry skill/skill/hook/agent in the target harness install dir and write a pass/fail report
+description: Use when verifying that a vibe install is intact or that CC and Codex are in sync — before a release, after changing skills/hooks/agents, or when a harness behaves inconsistently. Probes every shipped surface in the target install dir and writes a pass/fail report with STCV skill-quality verdicts.
 argument-hint: "[cc|codex]  (empty = current harness)"
 user-invocable: true
 ---
@@ -37,13 +37,13 @@ Execute the bundled implementation below with target harness: `$ARGUMENTS`
 - If `$ARGUMENTS` is empty, detect the current harness (CC vs Codex) and use that.
 - If the target install dir is missing, exit cleanly with guidance (not an error).
 
-See `skills/vibe.test/SKILL.md` for the probe spec and the report template.
+Probe spec: below. Report schema: `references/report-template.md`.
 
 ## Done Criteria
 
 - [ ] Runs without any external LLM call — file reads + vitest only
 - [ ] A single probe failure never halts the overall run
-- [ ] JSON report matches the template in SKILL.md exactly
+- [ ] JSON report matches `references/report-template.md` exactly
 - [ ] P1 failures auto-register via `/vibe.regress`
 
 ---
@@ -87,81 +87,100 @@ All probes are **structural or test-based** — no interactive command is ever a
 | 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`) |
+| agents | `<install>/agents/**/*.md` (**recursive**) | file readable · frontmatter parses · required fields (`name`, `description`) |
+| skill quality | `<install>/skills/*/SKILL.md` | STCV 4-axis quality check (below) |
 
 A probe's failure is captured in its `error` field; the overall run never halts because of one failure.
 
-## Report template
+**Agent discovery must recurse.** Agents live in subdirectories by capability group —
+`agents/ui/*.md` and `agents/event/*.md` are installed conditionally per stack/capability. A
+non-recursive `agents/*.md` glob silently skips 4 of the 11 shipped agents, so the report claims
+parity it never checked. Record each agent by its path-relative name (`ui/design-reviewer`, not
+`design-reviewer`) so group membership stays visible in the report.
+
+### Skill quality (STCV)
 
-Written to `~/.vibe/test-reports/<YYYYMMDD-HHmm>-<harness>.{json,md}`. Exact schema:
+Structural checks only — line counts, directory existence, and section/word presence. **No LLM call**, consistent with the rest of the probes.
 
-### JSON
+| Axis | Check | Verdict |
+|---|---|---|
+| **S**cope | SKILL.md line count | ≤250 `pass` · 251–400 `warn` · >400 `fail` |
+| **C**ontext | body fenced code blocks ≥3 **and** no sibling `references/` dir | `warn` |
+| **T**rigger | `description` contains no activation-condition wording | `warn` |
+| **V**erify | no `Done Criteria` / `완료 기준` section in body | `warn` |
+
+**WHY each axis** — a skill is loaded whole into context on invocation, so an oversized SKILL.md
+spends the caller's context on material most invocations never need (Scope), and material that
+*is* only needed sometimes belongs behind a reference pointer (Context). `description` is the
+only signal `/vibe` Catch-all routing sees, so a purely descriptive one cannot be matched against
+a request (Trigger). And a skill with no stated completion test cannot be judged by the
+deterministic JUDGE step in `vibe/rules/loop-contract.md` (Verify).
+
+Axis detail:
+
+- **Scope** — count lines of `SKILL.md` verbatim (frontmatter included). The `fail` threshold is
+  deliberately above the `warn` threshold so that a skill drifting past 250 gets flagged long
+  before it breaks the build.
+- **Context** — count ` ``` ` fence *pairs* in the body. Three or more blocks with no `references/`
+  sibling directory means templates/examples/schemas are inlined where they should be extracted.
+  A skill with a `references/` dir passes regardless of fence count.
+- **A `references/` dir is not by itself a pass.** The context axis asks whether *conditional bulk
+  still sits inline*, not whether the directory exists — a 6-line reference file next to a 150-line
+  body clears the directory check while saving nothing. Judge the body: does it still inline a
+  block that most invocations skip? If yes, the axis warns regardless of the sibling directory.
+- **Safety and correctness rules never move behind a conditional load.** "Read
+  `references/protected-branches.md` if the branch is protected" is circular — deciding whether the
+  rule applies requires already knowing the rule. Anything the skill must obey *before* it can
+  classify the situation (destructive-operation guards, push/force-push prohibitions, data-loss
+  boundaries) stays inline. Only material that is *looked up after* the situation is known is
+  eligible.
+- **Anti-gaming (scope + context)** — a reference only saves context if it is **conditionally**
+  loaded. If the body instructs reading a reference unconditionally ("read `references/x.md`" with
+  no condition attached), nothing was saved: the invocation now costs body + reference *plus* an
+  extra round trip. Emit `scope: warn` with `"unconditional reference load"` when a skill's body is
+  small but a reference it always loads is large — regardless of the line count passing. **A stub
+  that forwards to a wholesale copy of its own former body is a regression, not a split.** The test
+  for a real split: name a plausible invocation that never needs the reference. If none exists, the
+  material belongs inline.
+- **Trigger** — the `description` must contain at least one activation-condition marker:
+  `when` / `if` / `use for` / `use when` / `…할 때` / `…하면` / `…인 경우` / `…이 있고`.
+  A bare capability restatement (`"Execute implementation from SPEC"`) has none and warns.
+- **Verify** — case-insensitive search for a heading containing `Done Criteria` or `완료 기준`.
+
+**Alias exemption**: a merged skill leaves behind a thin alias — a body of ≤20 lines whose only
+content is a pointer at the canonical skill. Aliases are exempt from the **context** and **verify**
+axes: they hold no templates to extract and no work to complete, so demanding `references/` or a
+`Done Criteria` section would only pad them back into the duplication the merge removed. **Scope and
+trigger still apply** — an alias must stay small and must still say when it fires. Mark exempt axes
+`"n/a"` in the `quality` object rather than `"pass"`, so the report never claims a check it skipped.
+
+**Severity policy**: only `scope` can produce `fail`. The other three axes emit `warn` in this
+release — `verify` in particular currently applies to 17 of the 22 entry skills, and promoting it
+to `fail` before that backlog clears would ship a self-test that fails on vibe's own package and
+flood `/vibe.regress` with P1s. Promote `verify` to `fail` for entry skills in a later release
+once those skills carry Done Criteria.
+
+Quality findings attach to the existing skill entry rather than creating a second entry per skill:
 
 ```json
-{
-  "harness": "cc",
-  "timestamp": "2026-04-16T18:30:00+09:00",
-  "vibe_version": "2.9.24",
-  "install_dir": "/Users/grove/.claude",
-  "summary": {
-    "total": 42,
-    "passed": 40,
-    "failed": 2
-  },
-  "probes": {
-    "entrySkills": [
-      { "name": "vibe.spec", "status": "pass" },
-      { "name": "vibe.test", "status": "pass" }
-    ],
-    "skills": [
-      { "name": "test", "status": "pass" },
-      { "name": "spec", "status": "fail", "error": "invocation includes `auto` but triggers array is empty" }
-    ],
-    "hooks": [
-      { "name": "pre-tool-guard", "status": "pass", "tests": "38/38" },
-      { "name": "scope-guard", "status": "pass", "tests": "12/12" }
-    ],
-    "agents": [
-      { "name": "implementer", "status": "pass" }
-    ]
-  },
-  "failed": [
-    {
-      "category": "skills",
-      "name": "spec",
-      "error": "frontmatter: triggers array is empty"
-    }
-  ]
-}
+{ "name": "vibe.run", "status": "fail",
+  "quality": { "scope": "fail", "context": "warn", "trigger": "warn", "verify": "warn" },
+  "error": "scope: 845 lines (>400)" }
 ```
 
-- `status`: `"pass"` | `"fail"`
-- `error`: required when `status === "fail"`, omitted otherwise
-- `tests` (hooks only): short `passed/total` string from the vitest JSON reporter
-- `failed[]`: flat list of every failure for quick scanning — mirrors entries found across `probes.*`
-
-### Markdown
+A skill's `status` is the worst verdict across its structural check and all four axes
+(`fail` > `warn` > `pass`).
 
-```markdown
-# vibe self-test · cc · 2026-04-16 18:30
-
-**Version**: 2.9.24   **Install**: /Users/grove/.claude
-
-| Category | Pass | Fail |
-|---|---:|---:|
-| entry skills | 15 | 0 |
-| skills   | 17 | 1 |
-| hooks    |  6 | 0 |
-| agents   |  3 | 1 |
-| **total**| **41** | **2** |
+## Report template
 
-## Failures
+JSON 스키마와 Markdown 출력 포맷의 전문(필드·타입·명명·Warnings/STCV 섹션):
+`references/report-template.md`
 
-- **skills / spec** — invocation includes `auto` but triggers array is empty
-- **agents / implementer** — agent file not found
-```
+핵심 계약만 요약:
 
-If `failed` is empty, replace the Failures section with `_All probes passed._`.
+- `status`: `"pass"` | `"warn"` | `"fail"` — 구조 검사와 STCV 4축 중 **최악** 판정
+- `quality` (skills 만): `{ scope, context, trigger, verify }`, 각 `pass`/`warn`/`fail`/`n/a`
+- `failed[]` 는 `fail` 만, `warned[]` 는 `warn` 만. **`warned[]` 는 절대 `vibe.regress` 로 넘기지 않는다**
 
 ## Steps
 
@@ -169,11 +188,12 @@ If `failed` is empty, replace the Failures section with `_All probes passed._`.
 2. **Resolve install dir**: `cc` → `~/.claude`, `codex` → `~/.codex`. If missing → print guidance + exit.
 3. **Read `vibe_version`** from `package.json` in the current repo.
 4. **Walk each category**, run its check, append `{ name, status, error? }` to `probes.<category>`.
-5. **Compute** `summary` counts and the flat `failed[]` list.
-6. **Ensure** `~/.vibe/test-reports/` exists (`mkdir -p`, dir mode `0o700` — consistent with `~/.vibe/config.json`).
-7. **Write** `<ts>-<harness>.json` and `<ts>-<harness>.md`.
-8. **Print** the Markdown summary to the console.
-9. **If `summary.failed > 0`**, load skill `vibe.regress` with `subcommand: register --from-test` and pass the failed entries. P1 = any probe with `status: fail`.
+5. **Run the STCV check** on every `skills` entry, attach `quality`, and fold the worst axis verdict into that entry's `status`.
+6. **Compute** `summary` counts and the flat `failed[]` / `warned[]` lists.
+7. **Ensure** `~/.vibe/test-reports/` exists (`mkdir -p`, dir mode `0o700` — consistent with `~/.vibe/config.json`).
+8. **Write** `<ts>-<harness>.json` and `<ts>-<harness>.md`.
+9. **Print** the Markdown summary to the console.
+10. **If `summary.failed > 0`**, load skill `vibe.regress` with `subcommand: register --from-test` and pass the failed entries. P1 = any probe with `status: fail`. **`warned[]` is never passed to regress** — warnings are advisory and must not create P1 backlog.
 
 ## Done Criteria
 
@@ -184,4 +204,7 @@ If `failed` is empty, replace the Failures section with `_All probes passed._`.
 - [ ] 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`
+- [ ] `warned.length > 0` alone never invokes `vibe.regress`
+- [ ] Every `skills` entry carries a `quality` object with all four STCV axes
+- [ ] Agent discovery recurses — `agents/ui/*` and `agents/event/*` appear in the report under their group-qualified names
 - [ ] Entry skills are verified as user-invocable skill surfaces, not deprecated command files

package/skills/vibe.test/references/report-template.md

@@ -0,0 +1,101 @@
+# vibe.test — Report Template (JSON + Markdown)
+
+> vibe.test SKILL.md 의 Report 섹션에서 참조. `~/.vibe/test-reports/<ts>-<harness>.{json,md}` 의 정확한 스키마.
+
+## Report template
+
+Written to `~/.vibe/test-reports/<YYYYMMDD-HHmm>-<harness>.{json,md}`. Exact schema:
+
+### JSON
+
+```json
+{
+  "harness": "cc",
+  "timestamp": "2026-04-16T18:30:00+09:00",
+  "vibe_version": "2.9.24",
+  "install_dir": "/Users/grove/.claude",
+  "summary": {
+    "total": 42,
+    "passed": 38,
+    "warned": 2,
+    "failed": 2
+  },
+  "probes": {
+    "entrySkills": [
+      { "name": "vibe.spec", "status": "pass" },
+      { "name": "vibe.test", "status": "pass" }
+    ],
+    "skills": [
+      { "name": "vibe.verify", "status": "pass",
+        "quality": { "scope": "pass", "context": "pass", "trigger": "pass", "verify": "pass" } },
+      { "name": "vibe.run", "status": "fail",
+        "quality": { "scope": "fail", "context": "warn", "trigger": "warn", "verify": "warn" },
+        "error": "scope: 845 lines (>400)" },
+      { "name": "spec", "status": "fail", "error": "invocation includes `auto` but triggers array is empty" }
+    ],
+    "hooks": [
+      { "name": "pre-tool-guard", "status": "pass", "tests": "38/38" },
+      { "name": "scope-guard", "status": "pass", "tests": "12/12" }
+    ],
+    "agents": [
+      { "name": "implementer", "status": "pass" },
+      { "name": "ui/design-reviewer", "status": "pass" },
+      { "name": "event/event-planner", "status": "pass" }
+    ]
+  },
+  "failed": [
+    {
+      "category": "skills",
+      "name": "spec",
+      "error": "frontmatter: triggers array is empty"
+    }
+  ],
+  "warned": [
+    {
+      "category": "skills",
+      "name": "vibe.docs",
+      "error": "verify: no Done Criteria section"
+    }
+  ]
+}
+```
+
+- `status`: `"pass"` | `"warn"` | `"fail"` — worst verdict across the entry's structural check and its four quality axes
+- `error`: required when `status` is `"warn"` or `"fail"`, omitted otherwise. Prefix with the axis (`scope:`, `context:`, `trigger:`, `verify:`) when the verdict comes from a quality axis
+- `quality` (skills only): per-axis verdicts — `{ scope, context, trigger, verify }`, each `"pass"` | `"warn"` | `"fail"` | `"n/a"` (alias-exempt axis)
+- `tests` (hooks only): short `passed/total` string from the vitest JSON reporter
+- `failed[]`: flat list of every `fail` for quick scanning — mirrors entries found across `probes.*`
+- `warned[]`: same shape as `failed[]`, for `warn` entries. **Advisory only — never feeds `/vibe.regress`**
+
+### Markdown
+
+```markdown
+# vibe self-test · cc · 2026-04-16 18:30
+
+**Version**: 2.9.24   **Install**: /Users/grove/.claude
+
+| Category | Pass | Warn | Fail |
+|---|---:|---:|---:|
+| entry skills | 15 | 0 | 0 |
+| skills   | 15 | 2 | 1 |
+| hooks    |  6 | 0 | 0 |
+| agents   |  3 | 0 | 1 |
+| **total**| **39** | **2** | **2** |
+
+## Skill quality (STCV)
+
+| Skill | Scope | Context | Trigger | Verify |
+|---|---|---|---|---|
+| vibe.run | ❌ 845 | ⚠️ | ⚠️ | ⚠️ |
+| vibe.docs | ⚠️ 292 | ✅ | ✅ | ⚠️ |
+
+Only skills with at least one non-pass axis are listed. If every skill passes all four axes,
+replace this section with `_All skills pass STCV._`
+
+## Failures
+
+- **skills / spec** — invocation includes `auto` but triggers array is empty
+- **agents / implementer** — agent file not found
+```
+
+If `failed` is empty, replace the Failures section with `_All probes passed._`.

package/skills/vibe.tool-fallback/SKILL.md

@@ -2,7 +2,7 @@
 name: vibe.tool-fallback
 invocation: [auto]
 tier: optional
-description: "Tool failure fallback strategies with circuit breaker. Auto-activates on API errors, search failures, timeouts, 429, 5xx, overloaded errors."
+description: "Use when a tool or provider returns API errors, search failures, timeouts, 429, 5xx, or overload responses and a circuit-breaker fallback is required."
 triggers: [search failure, 429, 5xx, overloaded, fallback, circuit breaker, rate limit]
 priority: 60
 ---
@@ -33,32 +33,11 @@ priority: 60
 
 ## Decision Trees
 
-### Web Search Fails
+Select exactly one conditional reference for the failing capability; unrelated failures do not load it:
 
-```
-Web Search fails (429, 529, timeout)
-  → Check circuit state
-  → OPEN? → Skip to alternative immediately
-  → CLOSED? → Try context7 for library docs
-  → Still fails? → Claude's built-in knowledge (last resort)
-```
-
-### External LLM Fails
-
-```
-VibeOrchestrator.smartRoute({ type, prompt })
-  → Primary LLM fails (429, 401, 5xx)
-  → Skip to secondary LLM (no retry on rate limit)
-  → Secondary fails → Claude handles directly
-```
-
-### File/Code Not Found
-
-```
-Glob fails → Expand pattern: *.ts → **/*.ts → **/*
-  → Use Grep for content-based search
-  → Check git log for file history
-```
+- Web search failure → `references/web-search.md`
+- External LLM failure → `references/external-llm.md`
+- File/code lookup failure → `references/file-lookup.md`
 
 ## Error Response Actions
 
@@ -83,13 +62,12 @@ Request → Check circuit
                       └─ Use alternative
 ```
 
-## LLM Priority by Task Type
+## Model Selection
 
-| Task Type | Primary → Secondary → Fallback |
-|-----------|-------------------------------|
-| architecture, debugging | GPT → Antigravity → Claude |
-| uiux, code-analysis | Antigravity → GPT → Claude |
-| code-gen, general | Claude only |
+Inherit the active session model by default. Use the configured provider/model
+SSOT for external fallbacks instead of a hardcoded provider order. The only
+permitted override is an `opus` tier for an explicitly requested deep
+architecture review; if unavailable, keep the inherited model.
 
 ## Principles
 

package/skills/vibe.tool-fallback/references/external-llm.md

@@ -0,0 +1,10 @@
+# External LLM Fallback
+
+Load only for external LLM failures.
+
+```
+VibeOrchestrator.smartRoute({ type, prompt })
+  → Primary LLM fails (429, 401, 5xx)
+  → Skip to secondary LLM (no retry on rate limit)
+  → Secondary fails → active harness handles directly
+```

package/skills/vibe.tool-fallback/references/file-lookup.md

@@ -0,0 +1,9 @@
+# File and Code Lookup Fallback
+
+Load only for file or code lookup failures.
+
+```
+File-pattern search fails → Expand pattern: *.ts → **/*.ts → **/*
+  → Use content-based text search
+  → Check git log for file history
+```

package/skills/vibe.tool-fallback/references/web-search.md

@@ -0,0 +1,11 @@
+# Web Search Fallback
+
+Load only for web search failures.
+
+```
+Web Search fails (429, 529, timeout)
+  → Check circuit state
+  → OPEN? → Skip to alternative immediately
+  → CLOSED? → Try Context7 for library docs
+  → Still fails? → active harness knowledge (last resort, label uncertainty)
+```