Analyzes token spend across agent runs, separating cache-read from generated output costs.
npx skills add https://github.com/QSchlegel/session-viz --skill qcost-qschlegel
Please help me install this Agent Skill. Skill: qcost Source: https://github.com/QSchlegel/session-viz/tree/main/plugins/session-viz/skills/qcost Command: npx skills add https://github.com/QSchlegel/session-viz --skill qcost-qschlegel
node ${CLAUDE_PLUGIN_ROOT}/scripts/runs.mjs --cost to break down cache-read versus output tokens across Claude Code, Codex, and Cursor runs. - Sourced cost quoting: Fetches a dated rate card from the session-viz prices endpoint and only quotes money when the card is fresh, labelled with fetchedAt and source. - Agent family comparison: Surfaces the spread in cache-read per run between agent families, pointing to the editable prompt file responsible for the widest one. - Use Case: A user asks why their Claude Code bill is high; the skill shows that output is under half a percent of tokens, identifies which agent family replays the most context per run, and quotes an API-equivalent cost only if a fresh rate card exists. ## Quick Start Ask the assistant to run /qcost or ask where your token spend is going, and it will print the cache-read versus output breakdown for your recent agent runs.High-intent search queries and answers about installing and using this skill.
Run /qcost, which executes the local runs.mjs --cost report over your transcripts. It breaks down cache-read versus output tokens per agent family, showing that cache-read typically dominates total spend.
The dominant cost is cache-read: context replayed to the model on every turn, which never appears in per-session views. Output text is usually under half a percent of tokens, so reducing replayed context per run is what lowers cost.
Yes, the cost report reads only local transcripts and prints to stdout. Without network access it simply cannot fetch the pricing rate card, so it reports token counts without any currency figures.
Dollar figures are quoted only from a fetched, dated rate card and are labelled as API-equivalent cost, not actual subscription spend. If the card is missing or stale, no money is quoted at all.
The report cannot distinguish 5-minute from 1-hour cache write TTL pricing, and totals may exclude harnesses with missing or partial token data. A 'not in these numbers' block prints whenever coverage is incomplete.
tz5514Audit agent session transcripts for compliance and reasoning gaps.
leogurjaApplies consent-gated fixes to ranked token sinks from a Caveman learn report.
4lermanAnalyze AI API spending and implement model routing and prompt caching strategies.
ByronWilliamsCPATrack Claude Code token usage across local JSONL transcripts.