Costeamigo

Costeamigo — Historical Token Consumption Report

You generate a comprehensive, well-organized report of historical token consumption.

Step 1: Choose platform

Check $ARGUMENTS:

• If $ARGUMENTS is all → report all platforms combined
• If $ARGUMENTS is openclaw → report OpenClaw only
• If $ARGUMENTS is claude → report Claude Code only
• If $ARGUMENTS is codex → report Codex CLI only
• If $ARGUMENTS is empty or unrecognized → ask the user to choose:

Use AskUserQuestion with this format:

Choose a platform to analyze:

A) all     — Combined report (OpenClaw + Claude Code + Codex)
B) openclaw — OpenClaw sessions
C) claude   — Claude Code sessions
D) codex    — Codex CLI sessions

Enter A/B/C/D or platform name:

Map the user's response:

• A / all → all
• B / openclaw → openclaw
• C / claude → claude
• D / codex → codex

Step 2: Generate report data

Run the report script with the chosen source filter:

bash "SCRIPT_DIR/scripts/report.sh" --source <chosen_platform>

(Replace SCRIPT_DIR with the directory where this SKILL.md file lives.)

Valid --source values: all, openclaw, claude-code, codex

Note: user says "claude" but the script expects "claude-code". Map accordingly:

• claude → --source claude-code
• openclaw → --source openclaw
• codex → --source codex
• all → --source all

The script outputs JSON with these sections:

• overview: total tasks, tokens, cost, time range, sources
• by_source: breakdown per platform with models used
• by_skill: aggregation per skill
• by_model: aggregation per model
• by_tool: tool usage ranking
• reasoning_vs_tools: reasoning vs tool-invocation split
• top_tasks_by_tokens: top 10 most token-heavy tasks
• all_tasks: every task with full metrics

Step 3: Analyze and present

Using the JSON data, produce a clear, layered report:

Layer 1: Overview

• Total tasks, total tokens, total cost
• Time range covered
• Platform(s) analyzed

Layer 1.5: By Source Platform (only in all mode)

• Show each platform with: task count, tokens, models used
• Estimate cost for platforms without cost data using model pricing:
  • Claude Opus 4.6: $15/MTok input, $75/MTok output, $1.50/MTok cache read, $18.75/MTok cache write
  • Claude Sonnet: $3/MTok input, $15/MTok output
  • Claude Haiku: $0.80/MTok input, $4/MTok output
  • GPT-5.2-codex: $1.75/MTok input, $14/MTok output, $0.175/MTok cache read
  • GPT-5.4: $2.50/MTok input, $15/MTok output

Layer 2: By Category

By Skill:

• List each skill with: invocation count, total tokens, total cost, avg cost
• "(conversation)" = non-skill direct chat tasks

By Model:

• List each model with: task count, tokens (input/output/cache split), cost

By Tool:

• Rank tools by total call count

Layer 3: Efficiency Analysis

Reasoning vs Tool Invocation:

• Show the percentage split
• high reasoning % = mostly thinking; high tool % = lots of file I/O

Top Expensive Tasks:

• List top 5 most costly tasks
• Show: task prompt, tokens, cost, tools used

Layer 4: Insights (LLM analysis)

2-3 brief, actionable insights based on the data.

Formatting

• Use tables for structured data
• Use bold for key numbers
• Keep it scannable
• Use Chinese if the user communicates in Chinese, English otherwise