Summary

Add LRU cache with disk persistence and retry with exponential backoff to websearch, codesearch, and webfetch tools to improve performance and reliability.

Problem

The web tools currently:

1. No caching - Every identical query hits the network, even for repeated queries seconds apart
2. No retry - Transient network errors or rate limits cause immediate failure
3. Cold starts - Every OpenCode restart starts fresh with no cached data
4. No observability - No way to see if caching is working

Solution

1. LRU Cache with Disk Persistence

New utility at packages/opencode/src/util/cache.ts:

• In-memory LRU cache for fast lookups (<5ms)
• Disk persistence to ~/.cache/opencode/{namespace}/ for warm restarts
• TTL-based expiration for freshness
• SHA256 key hashing for safe filenames
• INFO-level logging for cache hits/misses (observability)
• Hit/miss counters with hit rate calculation

2. Retry with Exponential Backoff

Extended packages/util/src/retry.ts with new exports:

• isRetryableError() - Combines all retryable error checks
• isRateLimitError() - Detects 429 errors
• isServerError() - Detects 5xx errors

Configuration: 3 attempts, 1s → 2s → 4s backoff, max 10s delay

Retries: Network errors, rate limits (429), server errors (5xx)
Does NOT retry: Client errors (4xx) - those indicate request issues

Performance Impact

Observability

Cache operations are logged at INFO level:

INFO service=cache namespace=websearch event=hit source=memory
INFO service=cache namespace=websearch event=miss

Stats available via cache.stats():

{
  namespace: "websearch",
  memorySize: 42,
  maxSize: 500,
  hits: 156,
  misses: 23,
  hitRate: 0.87
}

Changes

Testing

• 17 new unit tests covering:
  • Cache behavior (LRU, TTL, eviction)
  • Hit/miss counters and hit rate
  • Counter reset on clear
  • Expired entries counting as misses
  • Retry logic with exponential backoff
• All existing tests pass
• Manual testing confirms cache hits on repeated queries

Breaking Changes

None. Existing behavior preserved, just faster and more reliable.