version: "1.0.0"

Skill: my-savings

Show the opt-in ClaudeRunway savings tracker: an estimate of context tokens avoided by local compression this session, plus this project's history. This includes both the local-compress MCP server's tools (compress_file, compress_command_output, fetch_url) and the compress_bash_output.py hook's own compressions of Bash/Grep/Glob/WebFetch/WebSearch output — both contribute to the same session total.

Optional argument: detail (e.g. /my-savings detail), trend (e.g. /my-savings trend, /my-savings trend day). Omit it for the simple view.

Important — this is an estimate, not a benchmark. This tracker computes a live, online estimate of a counterfactual ("what would this have cost without compression"), not a controlled A/B measurement. EVALUATION.md's tracks measure that rigorously by running a task twice; this skill reports what a heuristic estimator observed in real time. Don't present the two as equivalent.

When to use

Run this any time you want to check how much local compression has saved this session, or how this project trends over time. Requires CLAUDE_RUNWAY_TRACK_SAVINGS (accepted values: 1, true, or yes, case-insensitive) to be set in both .mcp.json's local-compress env block and exported in the shell environment that launches claude (hook entries in .claude/settings.json have no env field of their own, so they only ever see what the shell already has) — if it's off, the tools will say so.

Steps

1. Determine the view: if the argument is detail, use the detail view; if the argument is trend (optionally followed by day or week), use the trend view; otherwise use the simple view.

1. Simple view (no argument): call the savings_summary MCP tool (from the local-compress server) with no arguments — it defaults project to the current working directory's folder name. Print its return value to the user verbatim (it's pre-formatted text, not something to summarize or reformat).

1. Detail view (detail argument): call the savings_detail MCP tool the same way. Print its return value verbatim.

1. Trend view (trend argument): call the savings_trend MCP tool. If the user also supplied day or week after trend, pass that as the bucket parameter; otherwise the tool defaults to weekly buckets. Print its return value verbatim. Example: /my-savings trend day calls savings_trend(bucket="day").

1. If the tool reports tracking is off: relay that message plainly — don't try to work around it or guess a number. Tell the user both places CLAUDE_RUNWAY_TRACK_SAVINGS=1 needs to be set: their project's .mcp.json local-compress env block, and exported in the shell environment that launches claude (not .claude/settings.json — hook entries there have no env field of their own).

1. If the tool reports no session data yet: this is normal at the very start of a session, or in a project where nothing has been compressed yet. Don't treat it as an error — just relay what it says.

Notes

• This is read-only — it never writes, modifies, or deletes anything. (The perpetual SQLite rollup happens automatically at session end via a separate hook, not through this skill.)
• Only local-compression events are credited toward the headline number. Qdrant/qdrant-find activity is intentionally never counted here — its counterfactual (what Grep+Read would have cost) isn't observable, so crediting it would be a guess dressed up as a measurement. See the repo's EVALUATION.md.
• fetch_url calls are logged but never credited — its honest counterfactual is WebFetch's own already-compressed summary, which isn't observed. The tool's output will show fetch_url call counts separately from the credited savings figure.