Cache Efficiency

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, startedat, endedat, 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 cachereadtokens far exceeding cachewritetokens. When cache_reuse 70% strong, 40–70% moderate, > 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.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hoangsonww
• Source: hoangsonww/Claude-Code-Agent-Monitor
• License: MIT
• Homepage: https://hoangsonww.github.io/Claude-Code-Agent-Monitor/

Install and usage instructions live in the source repository linked above.