caveman-statslisted

# caveman-stats Reads the session transcript and reports actual token counts. Does NOT estimate what it can't read. ## Where the data is Claude Code writes a per-session JSONL transcript (assistant/user/tool messages), typically under the projects/transcripts dir of CLAUDE_DIR (path varies by build/version). Each assistant message carries a `usage` object with `input_tokens`, `output_tokens`, `cache_read_input_tokens`, `cache_creation_input_tokens`. ## Process 1. Locate the current session's JSONL (most-recently-modified transcript for this project, or the one whose session_id matches). 2. Sum `input_tokens`, `output_tokens`, and cache fields across assistant turns. 3. Report: total in/out, cache hit ratio, output tokens/turn. 4. Savings estimate (label it an ESTIMATE): caveman trims prose ~70%; apply only to the prose share of output (exclude code/tool args). State the assumption. ## If the fields aren't present Different builds expose usage differently. If you cannot find token fields, SAY SO and report what is available — never fabricate numbers. ## Output (caveman) ``` tokens: in <N> / out <N> · cache-read <N> (<P>%) out/turn: <N> over <T> turns est caveman saving: ~<P>% of prose output (assumption: prose≈<P> of out) ```