hoangsonww Cache Efficiency
Diagnose whether prompt caching is actually saving money, and where it is not.
Input
The user provides: $ARGUMENTS
This may be: empty (analyze the whole fleet), "today" / "this week" / a date range, a session ID to scope the analysis, or a target like "hit rate > 80%". When empty, analyze all data from /api/analytics.
Data Sources
| Endpoint | Returns |
|---|---|
GET /api/analytics |
tokens.total_input, tokens.total_output, tokens.total_cache_read, tokens.total_cache_write (baselines pre-summed), plus daily_sessions |
GET /api/sessions?limit=200 |
Session list — each has model, cwd, started_at, ended_at, inline cost, metadata (JSON: usage_extras with cache token detail) |
GET /api/sessions/{id} |
Full session detail with nested agents and events, for drill-down on a flagged session |
GET /api/pricing/cost |
{ total_cost, breakdown: [{ model, input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, cost, matched_rule }] } — used to price cache read vs write spend |
How cache economics work
cache_hit_rate = total_cache_read / (total_cache_read + total_input)
cache_reuse = total_cache_read / total_cache_write
cache_read_cost = (cache_read_tokens / 1M) × cache_read_per_mtok
cache_write_cost = (cache_write_tokens / 1M) × cache_write_per_mtok
Cache writes cost more per token than cache reads (e.g. Sonnet $3.75 write vs $0.30 read per Mtok), and writes are billed even if the cached block is never reused. The payoff only arrives on subsequent reads — so a healthy fleet shows cache_read_tokens far exceeding cache_write_tokens. When cache_reuse < 1, you are paying to cache context you barely re-read.
Token counts are effective totals = current + baseline (baselines preserve pre-compaction tokens).
Report Sections
1. Fleet Cache Hit Rate
From /api/analytics: compute cache_hit_rate × 100. State raw total_cache_read and total_input. Benchmark: >70% strong, 40–70% moderate, <40% weak prompt-cache utilization.
2. Write vs Read Reuse
Compute cache_reuse = total_cache_read / total_cache_write. Show both token counts. Flag if reuse < 1 (writing more cache than is ever read back).
3. Cache Spend Split
From /api/pricing/cost breakdown, sum cache_read_cost and cache_write_cost across all models. Show the dollar split and what fraction of total cost is cache-write overhead vs cache-read savings.
4. Sessions With Poor Reuse
From /api/sessions?limit=200, parse metadata.usage_extras for per-session cache read/write where available; rank sessions by lowest read/write reuse (and by cache_write-heavy cost). List the worst 10 with model, cost, and reuse ratio. Use /api/sessions/{id} to drill into any single flagged session.
5. Recommendations
- Sessions where
cache_write >> cache_read: short or one-shot sessions rarely recoup cache writes — note them. - Stable, repeated context (system prompts, large files) should be cached once and reused; high churn defeats caching.
- Estimate the dollar impact of raising the hit rate to the next benchmark tier.
Output
Structured Markdown with tables. Currency as USD to 4 decimal places; rates as $/Mtok; percentages with ▲/▼ for any trend. Token counts with thousands separators.