cost-summary
Skill2 starsChanged yesterdayMITView on GitHub
Use to inspect Anthropic API spend over a recent window — segments the `cost_log` table in `data/eyas.db` by surface (chat / triage routes / dashboards / scheduler) and by model, prints USD spend + % share + cache hit rate. Optional `--since <date>` window. Run before each QA cycle to confirm chat % stays near the 68% baseline and no surface has drifted on iterations or output cost. Trigger phrases: "what did the last week cost", "cost summary", "anthropic spend breakdown", "is chat still 68%".
In this repository8 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- audit-remediationsSkill
- audit-write-toolsSkill
- dev-serverSkill
- patch-contextSkill
- validate-bundleSkill
- verify-featureSkill
- verify-patchSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 62
After bumping a system prompt — re-run, confirm cache hit % didn't crater.
- Distinctive: line 63
Before a demo cycle — confirm spend is on a sustainable trend, not a 5x spike from an iteration loop bug.
- Distinctive: line 64
After a model swap — confirm the model bucket actually moved.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
cost-summary
Turns the append-only cost_log table in data/eyas.db into a one-shot view of where Anthropic spend is going. Designed to be run before each QA cycle and after large prompt changes — both the times you want to know whether you've drifted.
Run
bash .claude/skills/cost-summary/run.sh # full log
bash .claude/skills/cost-summary/run.sh --since 7d # last 7 days
bash .claude/skills/cost-summary/run.sh --since 24h
bash .claude/skills/cost-summary/run.sh --since 2026-05-01Output
Cost summary · 2026-04-26 → 2026-05-07 · 1,237 calls
By surface (USD) spend % calls ⌀ iter cache hit %
chat $42.18 68.4 412 7.3 82.1
triage-detections $5.91 9.6 98 4.1 71.4
briefing $4.02 6.5 31 3.0 88.2
…
TOTAL $61.66 100.0 1,237
By model
claude-sonnet-5-5 $58.21 94.4
claude-opus-5-5 $3.45 5.6
Cost shape (USD)
output $7.40 12.0% ← writing answers
cache write $19.83 32.2% ← first turn / new tools block
cache read $32.11 52.1% ← cached prefix replay
input (uncached) $2.32 3.8%
Headline ratios
output / total 12.0% ← higher = LLM is working hard, lower = cache + setup heavy
cache read / total 52.1% ← higher = better cache reuse
cache hit % 80.7% ← cache_read / (cache_read + cache_write + input)What it tells you
- Chat % of total. Baseline is ~68%. Drift much above means a route regressed (longer prompts, more iterations, less caching). Drift much below means chat usage dropped — usually a sign someone hit a bug.
- Output / total %. ~12% baseline. If this jumps, the model is generating more (longer responses, more tool-call retries). If it drops, the cache is doing more of the work.
- Cache hit %. Above 80% is healthy. Below 60% means the prompt prefix is changing call-to-call (rebuilds the cache instead of reusing) — usually a misplaced timestamp, randomized ordering, or a tools-block edit.
- ⌀ iter per surface. Average tool-call iterations. Chat baseline is ~7. Routes should be 3–5; spikes mean the agent is bouncing on a tool error or re-trying.
Limits
- Prices come straight from
lib/cost/pricing.ts(longest-prefix match on model id), the same table the in-app /cost page uses. - A surface that doesn't appear in the log either wasn't called in the window or never logs (the read-only dashboard fetchers don't go through
logUsage()because they don't run the agent loop). - Doesn't cross-reference
data/perf-log.jsonl— that'snpm run perf-summary. Use the two together: cost surfaces what you're paying for; perf surfaces what's slow.
When this catches things
- After bumping a system prompt — re-run, confirm cache hit % didn't crater.
- Before a demo cycle — confirm spend is on a sustainable trend, not a 5x spike from an iteration loop bug.
- After a model swap — confirm the model bucket actually moved.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

