cc-waste

Scans your local Claude Code session logs and generates a single self-contained HTML report showing token usage, estimated cost, and likely token-waste patterns — repeated file reads, oversized tool results, large pasted content, low prompt-cache hit rates, and unused or rarely-used installed skills. With --fix, it disables the unused/rarely-used skills itself and the report shows exactly what it found and what it did about it.

100% local. It only reads files under your own ~/.claude/projects/ directory. It makes no network calls and sends nothing anywhere.

Early version. Token/cost numbers are heuristic estimates (not exact), and this hasn't been battle-tested against every Claude Code setup yet (Windows, huge logs, etc.). Feedback and bug reports are very welcome — open an issue.

Usage

npx cc-waste

This writes cc-waste-report.html in the current directory, prints a summary to the terminal, and opens the report in your default browser — no extra flag needed. If any installed skill looks unused, the terminal and the report both print the exact command to clean it up.

Options

-o, --out <file>     Output HTML path (default: ./cc-waste-report.html)
--dir <path>         Claude config dir to scan (default: $CLAUDE_CONFIG_DIR or ~/.claude)
--days <n>           Only include activity from the last N days
--project <substr>   Only include projects whose folder name contains this
--json               Also write a .json file with the raw aggregated data
--no-open            Don't auto-open the report in a browser (default: opens)
--fix                Disable unused/rarely-used skills automatically (see below)
-h, --help           Show help

Examples:

# Last 30 days only
npx cc-waste --days 30

# One project, with raw JSON alongside the HTML, no auto-open
npx cc-waste --project my-app --json --no-open

# Also disable the skills that turned out to be unused
npx cc-waste --fix

What counts as "waste"

The tool is heuristic, not exact — token counts are estimated at roughly 4 characters per token, and costs use approximate Anthropic list pricing by model family (see src/pricing.js). Treat every number as directional.

| Pattern | What it flags | |---|---| | Oversized tool result | A single tool call (Read, Bash, Grep, ...) returned a very large result. | | Repeated full-file read | The same file was read in full more than once in one session. | | Large pasted content | A user turn pasted a large block of text/code directly instead of referencing a file. | | Low prompt-cache hit rate | A multi-turn session had a low share of cached input tokens, usually from unstable context between turns. | | Unused skill | A skill installed under ~/.claude/skills was never invoked in any scanned session (needs ≥5 scanned sessions to flag). | | Rarely-used skill | A skill was invoked once or less across ≥10 scanned sessions. |

Every installed skill's name + description sits in the tool catalog every session whether or not it's ever used, so an unused skill has a real, ongoing token cost — the report estimates it the same way as the other findings.

Self-fixing unused skills (--fix)

Without --fix, the report only lists what it found — nothing on disk is touched. If it found any unused/rarely-used skills, both the terminal output and the report's "설치된 스킬 감사" section print the exact command to run next (your original --dir/--project/--days carried over, plus --fix), so you can just copy-paste it.

With --fix, cc-waste moves each unused/rarely-used skill's folder from ~/.claude/skills/<name> to ~/.claude/skills-disabled/<name>. This disables it without deleting it — Claude Code only discovers skills under skills/, so the moved folder simply stops showing up, and you can restore it any time with the exact mv command the report prints for it (also shown in the "설치된 스킬 감사" / "installed skills audit" section of the HTML). A skill that's still frequently used is never touched.

Skills wired into a Claude Code hook are never disabled, even by --fix. Some skills ship a hook (e.g. a Stop hook) referenced by file path in settings.json — moving the skill's folder away breaks that hook ("No such file or directory") even though nothing was deleted. Before disabling anything, cc-waste scans settings.json/settings.local.json (global and, for every project seen in the scanned logs, per-project) for any string that references a flagged skill's folder, and marks a match "보호됨 (hook 연결)" — reported like any other finding, but --fix skips it and leaves it alone.

How it works

1. src/find-logs.js locates every *.jsonl transcript under <claude home>/projects/*/.
2. src/parse-session.js streams each transcript line by line (no full-file loads), accumulating token usage per model and matching each tool call to its result to compute the waste findings above.
3. src/aggregate.js rolls sessions up into per-day, per-project, and per-model totals plus a ranked findings list.
4. src/skills-audit.js reads every installed skill's frontmatter under <claude home>/skills/*/SKILL.md, cross-references call counts collected in step 2 (from Skill tool calls), flags unused/rarely-used ones, checks each against settings.json hook references, and — only with --fix — moves the unprotected ones to skills-disabled/.
5. src/html-report.js renders it all into one static HTML file — no external scripts, stylesheets, or CDN dependencies, so it works fully offline.

No dependencies beyond Node's built-in modules (node >= 18).

Support

If this saved you some tokens, you can buy the author a coffee: ko-fi.com/slowwave. Completely optional — cc-waste stays free, local-only, and doesn't nag you about it beyond a quiet link in the report footer and npm install's funding notice.

License

MIT © slow.wave — see LICENSE.