Usage analysis with tokentaby

tokentaby logs every Claude Code turn in this repo (tokens, model, agent, task, branch, estimated cost) to .tokentaby/devs/<dev>/<YYYY-MM>.ndjson, plus the sessions developers ran here in other coding agents (each event's tool: claude-code, codex-app, codex-cli, opencode, gemini-cli, qwen-code, copilot-cli, kimi-code, hermes-agent, cursor), committed to git, so a pulled clone holds the whole team's usage. Query it with the bundled CLI — it prints JSON:

node "${CLAUDE_PLUGIN_ROOT}/scripts/tokentaby.js" query <view> [filters]

view	answers
summary	totals for a range, previous-period comparison, daily series, model / agent / tool share (byTool), top devs and tasks
periods	today / this week / this month, the current dev vs the team
devs	per developer: sessions, prompts, tokens by type, cost, favourite model, subagent share
dev <id>	one developer: timeline, model mix, tasks, recent prompts
models	per model: cost, tokens, requests, prompts answered, which devs / tasks / agents used it
agents	spend by agent type (main vs Explore, Plan, custom…) and subagent overhead per session
tasks	per task and per branch: cost, tokens, prompts, models, devs, active time
prompts	the prompt log (respects the repo's prompt mode); --q TEXT searches text, task id or hash
events	raw events, for anything the views do not cover

Filters: --from YYYY-MM-DD --to YYYY-MM-DD --dev ID --tool ID --model ID --task ID --agent TYPE --branch NAME --q TEXT. Add --prices current to re-cost history with today's pricing table instead of the price logged at the time. summary defaults to the current month; most other views default to all time — pass a range when the question names one.

Examples:

• "What did TKT-142 cost?" → query tasks --task TKT-142
• "Who used the most Opus this week?" → query devs --model claude-opus-5-5 --from <monday> --to <today>
• "How much do subagents add?" → query agents
• "Codex vs Claude Code this month?" → query summary and read byTool; one tool only → query summary --tool codex-cli

How to answer

• Always call costs estimated (API-equivalent). The team is on Claude subscriptions; these are list-price estimates for comparison and budgeting, not a bill.
• Tokens come in four kinds: input, output, cache write and cache read. Cache reads dominate raw token counts, so lead with estimated cost, and mention the token mix only when it matters.
• Usage not yet committed sits in this clone's queue, so only this developer's newest turns are visible before they commit. Teammates' usage arrives with git pull.
• If the CLI says tokentaby is not set up, tell the user to run /tokentaby:init (then /tokentaby:backfill for their history) and commit .tokentaby/.
• Never print raw prompt text beyond what the user asked for. In hash_only mode no text exists; say so instead of guessing.