Real usage, one operator: 2026-10-19 and the last 30 days

Downshift by Tiago Vilas Boas. Updated . Source on GitHub

What this page answers: what did Downshift do on a real working day in Claude Code, and what can and cannot be claimed from it? Source: downshift stats --days=1 --detail, downshift stats --days=30 --detail, downshift report --days=1 and --days=30, run on 2026-10-19 on the maintainer's machine. Aggregates only: no prompts, task text, paths or session ids.

Downshift by Tiago Vilas Boas · https://github.com/tiagovilasboas/downshift

The day

A long Claude Code session that built a landing page, restyled the Claude Code mods, wrote copy and ran audits. The parent agent delegated most of the work as long briefs to general-purpose subagents, so the tasks were natural, not mechanical.

Measure 2026-10-19
Subagent decisions 61
Downshifted (rewrite applied) 5 (8.2%)
Upshifted (rewrite applied) 0
Classified shift held by a guardrail, spawn unchanged 16 (26.2%)
Unchanged, model already right 27 (44.3%)
No known baseline 13 (21.3%)
Frontier rewrites on doubt 0
Complexity: medium / complex / trivial 70.5% / 24.6% / 4.9%
Observed child cost, catalog priced $50.45 (47 usage records)
Estimated net saving, list price $1.37 (about 2.7% of observed child cost)
Control-group spawns / with an end record 6 / 0

The last 30 days (same operator, several harnesses)

Measure 30 days
Subagent decisions 1,217
Downshifted / upshifted 199 (16.4%) / 61 (5.0%)
Held by a guardrail (safety-held) 243
Unchanged 255 (21.7%)
No known baseline 528 (44.9%)
Estimated net saving, list price $22.43 (saved $52.38, cost more $29.96)
Normalised units, routed against unrouted -1.9%
Control-group spawns / with an end record 31 / 0

What this supports

  • The hook decided on every delegated spawn without a model call: 61 decisions today, 1,217 in 30 days.
  • Doubt did not spend money: 0 frontier rewrites on doubt today; 16 classified moves were held by a guardrail today and 243 in 30 days.
  • On natural, design-heavy delegation the router mostly keeps the model: 5 of 61 downshifts today. This matches the 2026-10-15 trial (subagent-routing-ab-2026-10-15.md): savings concentrate on mechanical work.

What it does not support

  • Any saving percentage for general use. The $1.37 and $22.43 are per-decision list-price estimates, not measurements, and the 30-day dollar estimate and the normalised units disagree in sign, so the net effect is inside the noise.
  • A routed-versus-control comparison. downshift report needs 20 completed tasks per group; the control group had 0 completed (no end record) both today and over 30 days, so the measured comparison cannot be computed yet.

Defects this exposed

Defect Evidence Effect
Control-group spawns never get an end record 0 of 6 today, 0 of 31 in 30 days, while routed spawns have 43 and 214 downshift report can never compare routed with control
Custom agent types have no baseline 3 of 3 today, 72 of 72 in 30 days unknown (claude-code-guide 1 of 1) those spawns cannot be routed or priced

More docs