@su-record/vibe 2.6.14 → 2.6.16

package/skills/tool-fallback.md

@@ -1,190 +1,190 @@
----
-description: Tool failure fallback strategies with circuit breaker. Auto-activates on API errors, search failures, timeouts, 429, 5xx, overloaded errors.
----
-# Tool Fallback Strategies
-
-Guide for finding alternatives when tools fail to continue work.
-
-## Circuit Breaker Pattern
-
-Track tool failures and temporarily disable unreliable tools:
-
-```
-Tool State Machine:
-┌─────────┐   3 failures   ┌─────────┐   30s cooldown   ┌─────────────┐
-│ CLOSED  │ ─────────────→ │  OPEN   │ ───────────────→ │ HALF-OPEN   │
-│ (normal)│                │ (block) │                  │ (test 1 req)│
-└─────────┘                └─────────┘                  └─────────────┘
-     ↑                                                        │
-     │                      success                           │
-     └────────────────────────────────────────────────────────┘
-```
-
-| State | Behavior |
-|-------|----------|
-| **CLOSED** | Normal operation, count failures |
-| **OPEN** | Block all requests, use alternative immediately |
-| **HALF-OPEN** | Allow 1 test request after cooldown |
-
-**Per-tool tracking:**
-```
-GPT hook: failures=2, state=CLOSED
-Gemini hook: failures=0, state=CLOSED
-WebSearch: failures=3, state=OPEN (blocked until 14:32:00)
-context7: failures=1, state=CLOSED
-```
-
-## When Web Search Fails
-
-| Alternative | Method |
-|-------------|--------|
-| context7 plugin | `resolve-library-id` → `get-library-docs` |
-| VibeOrchestrator | `smartWebSearch()` - auto fallback chain |
-| Cached knowledge | Use Claude's built-in knowledge |
-
-```
-Web Search fails (429, 529, timeout)
-    ↓
-Check circuit breaker state
-    ↓
-If OPEN → Skip to alternative immediately
-    ↓
-If CLOSED → Try context7 for library docs
-    ↓
-If still fails: VibeOrchestrator.smartWebSearch()
-    ↓
-Last resort: Claude's built-in knowledge
-```
-
-## API Error Responses
-
-| Error | Cause | Response | Circuit Breaker |
-|-------|-------|----------|-----------------|
-| 429 | Rate Limit | Exponential backoff | +1 failure count |
-| 5xx | Server Error | Switch to alternative | +1 failure count |
-| 529 | Overloaded | Wait and retry | +1 failure count |
-| Timeout | Network | Split request or retry | +1 failure count |
-| 401/403 | Auth Error | Re-auth or alternative | Don't count (auth issue) |
-
-## When File/Code Not Found
-
-```
-Glob fails
-    ↓
-Expand pattern: *.ts → **/*.ts → **/*
-    ↓
-Use Grep for content-based search
-    ↓
-Check git log for history
-```
-
-## When External LLM Fails
-
-**Production (VibeOrchestrator):**
-
-```
-VibeOrchestrator.smartRoute({ type, prompt })
-    ↓
-LLM priority based on task type:
-  - architecture/debugging: GPT → Gemini → Claude
-  - uiux/code-analysis: Gemini → GPT → Claude
-  - code-gen: Claude only
-    ↓
-Auto fallback on primary LLM failure
-    ↓
-Claude handles directly when all fail
-```
-
-**Test/Debug Hooks (development only):**
-
-```
-gpt- or gpt. [question] fails
-    ↓
-Check circuit: If OPEN, skip to next
-    ↓
-Try gemini- or gemini. [question] (similar capability)
-    ↓
-Try context7 (for docs)
-    ↓
-Claude solves alone
-```
-
-> **Note:** `test-gpt`, `test-gemini` prefixes are for hook connectivity testing only.
-> For actual work, use VIBE commands (`/vibe.run`, `/vibe.spec`, etc.) and
-> VibeOrchestrator will automatically select the appropriate LLM.
-
-## Retry Strategy with Circuit Breaker
-
-```
-Request to tool
-    ↓
-Check circuit state
-    ↓
-┌─ OPEN? ──→ Use alternative immediately (no retry)
-│
-└─ CLOSED/HALF-OPEN? ──→ Try request
-                              ↓
-                         Success? ──→ Reset failure count
-                              ↓ No
-                         Retry with backoff:
-                           retry(1): wait 2s
-                           retry(2): wait 4s
-                           retry(3): wait 8s
-                              ↓
-                         All failed? ──→ +1 failure, check threshold
-                              ↓
-                         failures >= 3? ──→ OPEN circuit for 30s
-                              ↓
-                         Use alternative
-```
-
-## VibeOrchestrator Smart Routing
-
-VIBE commands (`/vibe.spec`, `/vibe.run`, etc.) use VibeOrchestrator internally.
-
-### LLM Priority by Task Type
-
-| Task Type       | Primary | Secondary | Fallback |
-|-----------------|---------|-----------|----------|
-| `architecture`  | GPT     | Gemini    | Claude   |
-| `debugging`     | GPT     | Gemini    | Claude   |
-| `uiux`          | Gemini  | GPT       | Claude   |
-| `code-analysis` | Gemini  | GPT       | Claude   |
-| `web-search`    | GPT     | Gemini    | Claude   |
-| `code-gen`      | Claude  | -         | -        |
-| `general`       | Claude  | -         | -        |
-
-### Auto Fallback Logic
-
-```
-smartRoute({ type: 'architecture', prompt })
-    ↓
-1. Try GPT (max 2 retries)
-    ↓ fails (429, 401, 5xx)
-2. Try Gemini (max 2 retries)
-    ↓ fails
-3. Claude handles directly (fallback message)
-```
-
-### Retry vs Immediate Switch
-
-| Error            | Action                 |
-|------------------|------------------------|
-| 429 Rate Limit   | Skip to next LLM       |
-| 401/403 Auth     | Skip to next LLM       |
-| Network Error    | Retry with backoff     |
-| 5xx Server Error | Retry then switch      |
-
-### Availability Cache
-
-- 5-minute TTL for LLM status cache
-- Auto-disable after 3 consecutive failures
-- Failed LLMs are skipped in subsequent requests
-
-## Principles
-
-1. **Never stop** - Always find an alternative
-2. **Before asking user** - Try alternatives first
-3. **Track failures** - Open circuit after 3 consecutive failures
-4. **Auto-recover** - Test after 30s cooldown
-5. **Fail fast** - Skip blocked tools immediately
+---
+description: Tool failure fallback strategies with circuit breaker. Auto-activates on API errors, search failures, timeouts, 429, 5xx, overloaded errors.
+---
+# Tool Fallback Strategies
+
+Guide for finding alternatives when tools fail to continue work.
+
+## Circuit Breaker Pattern
+
+Track tool failures and temporarily disable unreliable tools:
+
+```
+Tool State Machine:
+┌─────────┐   3 failures   ┌─────────┐   30s cooldown   ┌─────────────┐
+│ CLOSED  │ ─────────────→ │  OPEN   │ ───────────────→ │ HALF-OPEN   │
+│ (normal)│                │ (block) │                  │ (test 1 req)│
+└─────────┘                └─────────┘                  └─────────────┘
+     ↑                                                        │
+     │                      success                           │
+     └────────────────────────────────────────────────────────┘
+```
+
+| State | Behavior |
+|-------|----------|
+| **CLOSED** | Normal operation, count failures |
+| **OPEN** | Block all requests, use alternative immediately |
+| **HALF-OPEN** | Allow 1 test request after cooldown |
+
+**Per-tool tracking:**
+```
+GPT hook: failures=2, state=CLOSED
+Gemini hook: failures=0, state=CLOSED
+WebSearch: failures=3, state=OPEN (blocked until 14:32:00)
+context7: failures=1, state=CLOSED
+```
+
+## When Web Search Fails
+
+| Alternative | Method |
+|-------------|--------|
+| context7 plugin | `resolve-library-id` → `get-library-docs` |
+| VibeOrchestrator | `smartWebSearch()` - auto fallback chain |
+| Cached knowledge | Use Claude's built-in knowledge |
+
+```
+Web Search fails (429, 529, timeout)
+    ↓
+Check circuit breaker state
+    ↓
+If OPEN → Skip to alternative immediately
+    ↓
+If CLOSED → Try context7 for library docs
+    ↓
+If still fails: VibeOrchestrator.smartWebSearch()
+    ↓
+Last resort: Claude's built-in knowledge
+```
+
+## API Error Responses
+
+| Error | Cause | Response | Circuit Breaker |
+|-------|-------|----------|-----------------|
+| 429 | Rate Limit | Exponential backoff | +1 failure count |
+| 5xx | Server Error | Switch to alternative | +1 failure count |
+| 529 | Overloaded | Wait and retry | +1 failure count |
+| Timeout | Network | Split request or retry | +1 failure count |
+| 401/403 | Auth Error | Re-auth or alternative | Don't count (auth issue) |
+
+## When File/Code Not Found
+
+```
+Glob fails
+    ↓
+Expand pattern: *.ts → **/*.ts → **/*
+    ↓
+Use Grep for content-based search
+    ↓
+Check git log for history
+```
+
+## When External LLM Fails
+
+**Production (VibeOrchestrator):**
+
+```
+VibeOrchestrator.smartRoute({ type, prompt })
+    ↓
+LLM priority based on task type:
+  - architecture/debugging: GPT → Gemini → Claude
+  - uiux/code-analysis: Gemini → GPT → Claude
+  - code-gen: Claude only
+    ↓
+Auto fallback on primary LLM failure
+    ↓
+Claude handles directly when all fail
+```
+
+**Test/Debug Hooks (development only):**
+
+```
+gpt- or gpt. [question] fails
+    ↓
+Check circuit: If OPEN, skip to next
+    ↓
+Try gemini- or gemini. [question] (similar capability)
+    ↓
+Try context7 (for docs)
+    ↓
+Claude solves alone
+```
+
+> **Note:** `test-gpt`, `test-gemini` prefixes are for hook connectivity testing only.
+> For actual work, use VIBE commands (`/vibe.run`, `/vibe.spec`, etc.) and
+> VibeOrchestrator will automatically select the appropriate LLM.
+
+## Retry Strategy with Circuit Breaker
+
+```
+Request to tool
+    ↓
+Check circuit state
+    ↓
+┌─ OPEN? ──→ Use alternative immediately (no retry)
+│
+└─ CLOSED/HALF-OPEN? ──→ Try request
+                              ↓
+                         Success? ──→ Reset failure count
+                              ↓ No
+                         Retry with backoff:
+                           retry(1): wait 2s
+                           retry(2): wait 4s
+                           retry(3): wait 8s
+                              ↓
+                         All failed? ──→ +1 failure, check threshold
+                              ↓
+                         failures >= 3? ──→ OPEN circuit for 30s
+                              ↓
+                         Use alternative
+```
+
+## VibeOrchestrator Smart Routing
+
+VIBE commands (`/vibe.spec`, `/vibe.run`, etc.) use VibeOrchestrator internally.
+
+### LLM Priority by Task Type
+
+| Task Type       | Primary | Secondary | Fallback |
+|-----------------|---------|-----------|----------|
+| `architecture`  | GPT     | Gemini    | Claude   |
+| `debugging`     | GPT     | Gemini    | Claude   |
+| `uiux`          | Gemini  | GPT       | Claude   |
+| `code-analysis` | Gemini  | GPT       | Claude   |
+| `web-search`    | GPT     | Gemini    | Claude   |
+| `code-gen`      | Claude  | -         | -        |
+| `general`       | Claude  | -         | -        |
+
+### Auto Fallback Logic
+
+```
+smartRoute({ type: 'architecture', prompt })
+    ↓
+1. Try GPT (max 2 retries)
+    ↓ fails (429, 401, 5xx)
+2. Try Gemini (max 2 retries)
+    ↓ fails
+3. Claude handles directly (fallback message)
+```
+
+### Retry vs Immediate Switch
+
+| Error            | Action                 |
+|------------------|------------------------|
+| 429 Rate Limit   | Skip to next LLM       |
+| 401/403 Auth     | Skip to next LLM       |
+| Network Error    | Retry with backoff     |
+| 5xx Server Error | Retry then switch      |
+
+### Availability Cache
+
+- 5-minute TTL for LLM status cache
+- Auto-disable after 3 consecutive failures
+- Failed LLMs are skipped in subsequent requests
+
+## Principles
+
+1. **Never stop** - Always find an alternative
+2. **Before asking user** - Try alternatives first
+3. **Track failures** - Open circuit after 3 consecutive failures
+4. **Auto-recover** - Test after 30s cooldown
+5. **Fail fast** - Skip blocked tools immediately