bench — the proof behind the savings claim

python -m skills.bench.cli            # table + totals
python -m skills.bench.cli --json
python -m skills.bench.cli --save md  # timestamped report under .botte/reports/

Runs tasks.py's fixed corpus (17 prompts spanning trivial → hard reasoning) through the real auto_router.decide() — decision only, 0 tokens spent measuring — and compares the actual cost of what got picked against a documented baseline (every task sent straight to Tier.STANDARD, i.e. what a project with no routing at all pays). The delta is the number: token % saved, USD % saved, % that stayed local.

Why this exists

The README says "~65% (reported by users)" — a claim, not a proof. bench is versioned in the repo, runs offline, and gives the same shape of number every time: python -m skills.bench.cli in CI or before a release, diff the totals against the last run. It's also the evidence for [[preflight]]'s "prefer local" default and for pitching the belt to other agent frameworks (see the Hermes-Agent integration proposal in docs/plans/).

Reading the result

• local_pct — % of the corpus that resolved to a local backend (0 cloud tokens for those).
• totals.token_savings_pct / usd_savings_pct — belt cost vs. the fixed baseline.
• Caveat, stated plainly: the corpus prompts are short (they're meant to be reproducible and fast, not exhaustive), so the effort estimator's length-based signal under-weights them relative to real production prompts — which carry full stack traces, diffs, multi-file context. On a machine with a local backend running, this bench tends to read higher local% than a real session would. Treat it as a floor/sanity-check number, not a substitute for control_loop.analyze()'s real historical stats.

Related: [[auto_router]] (what's being measured), [[control_loop]] (the real, historical counterpart), [[cost_estimator]] (the cost model reused here).