What problem does it solve? Claude Code sessions silently accumulate token costs through oversized contexts, prompt cache rebuilds, unpinned subagent models, and large tool outputs, with no built-in way to see where the spend goes. This Skill reads local session transcripts and reports exactly which cost-equation terms drive your usage and what to change. ## Core Features & Use Cases - Weighted spend breakdown: Aggregates input, cache write, cache read, and output tokens per model across main and subagent transcripts, using Anthropic price ratios as relative weights (never absolute dollar amounts). - Anti-pattern flag detection: Automatically flags prompt-init overhead, context over 400K, prefix rewrite waste, model-switch rewrites, idle-expiry rewrites, large tool results, unpinned subagent models, subagent underuse, output-heavy sessions, high effort defaults, and MCP-heavy usage, each with metric, impact, and remediation. - Prefix rebuild cause analysis: Classifies every full cache rewrite by cause (session start, compaction, model switch, idle gaps) with event counts and token shares. - Use Case: After noticing a project burns tokens unusually fast, run the audit to discover that 40% of cache writes come from mid-session model switches and that Read results over 40KB dominate tool output, then pin subagent models and delegate bulk reads. ## Quick Start Ask the assistant to run a session cost audit on the current project and report the top token waste flags with recommended fixes.
