Spend · Token Diet

Token Diet

Is my token diet real? A partner (or you) claims a token saving. This view joins the signals RepoOps already produces into one verdict: real, at risk, unproven, or no diet yet. It adds no detection of its own.

Where to find it

  • Localhost: /token-diet.html?repo=<id>
  • API: GET /api/token-diet?repo=<id>
  • Keyboard: ⌘ K then token diet
  • Sidebar: Spend → Token Diet

What it does for you

One verdict, four honest states.The header reads real, at risk, unproven, or no diet yet. A saving that lost quality, or a bill that disagrees with your cache reuse, reads at risk, never real. A repo with no telemetry reads no diet yet.
Every signal shown, present or empty.Each signal RepoOps produces gets its own card: realized savings and the quality join, the Billing Guard cache-share reconciliation, the context-window composition, routing-tier coverage, subagent-delegation savings, the eval bars, the prompt-caching ROI, and model-downgrade verification. A signal with no backing data reads honestly empty, never a fabricated pass.
A verdict trend that builds over time.The verdict is recorded once per UTC day. The trend row shows recent days and this-week-vs-last, so a week of history builds up as you check back. Empty until the first day is recorded.

Use it well

Pick a repo with Claude Code telemetry, then read the header first. If it says at risk, open the cards to find which signal contradicts the claim: a failed eval bar, a negative caching ROI, or a cache-share divergence. The cache-share signal needs an Anthropic Admin key to reconcile the bill; without one it stays honestly empty rather than passing.

Read more