Fallback

Fallback

Graceful degradation when primary approach fails. If preferred agent/skill/tool fails, try alternates in priority order. Record which path succeeded for future routing decisions.

Purpose

Increase resilience. Instead of crashing on first failure, attempt a ladder of fallback strategies. Log outcomes for analytics and future routing optimization.

When to use

• Primary API rate-limited → try cheaper model or cached result
• Preferred tool unavailable (Playwright, Firecrawl) → shell equivalent or simpler approach
• Test framework broken → manual verification or other test harness
• Deployment target down → try alternate region or staging

Fallback Ladder Pattern

1. Try primary — Execute preferred approach.
2. On failure — Capture error, severity, and root cause.
3. Try secondary — Run alternate (lower cost, lower latency, or lower feature fidelity).
4. Try tertiary — Final fallback (minimal viable approach).
5. Escalate — If all fail, ask user or escalate to architect.
6. Record — Log which path succeeded + why for future routing.

Example Ladder

Task: "Extract text from PDF at https://example.com/doc.pdf"

Primary: Firecrawl + advanced extraction (costs $0.10, ~2s)
  → Fails: Rate limit / unavailable
  
Secondary: pypdf + manual page parse (costs $0, ~5s)
  → Succeeds: Extracts text (lower quality)
  → Log: "fallback-pdf-extraction:pypdf" for next time

---

Task: "Validate auth flow under load"

Primary: k6 load test (needs staging env)
  → Fails: Staging DB offline
  
Secondary: Locust (Python, lower overhead)
  → Succeeds: Simulates load
  → Log: "fallback-loadtest:locust" for future
  
Tertiary: Manual curl loop (minimal viable)
  → Would execute if Locust also failed

Invocation

fallback --task "Extract data" --primary "firecrawl" --secondary "pypdf" --tertiary "pdfplumber"

Or via skill trigger: "Try fallback for PDF extraction"

Logging Format

Each fallback attempt logs to ~/.claude/logs/fallback.jsonl:

{
  "timestamp": "2026-04-16T14:22:00Z",
  "task": "pdf-extraction",
  "primary": "firecrawl",
  "status": "failed",
  "error": "rate_limit_exceeded",
  "fallback_used": "pypdf",
  "result": "success",
  "cost_saved": 0.10,
  "latency_impact": "+3s"
}

References

• Smart model select: smart-model-select (similar fallback for model choice)
• Dispatch skill: dispatch (routes to alternates)
• Error handling patterns: best-practices/error-handling.md