Usage page doesn't disclose what the cost estimate excludes (Engine spend) or over-counts (cache)

[serge-ivo]

The gap

store/console/src/pages/Usage.tsx says:

Token usage and estimated cost across all your agents. Cost is estimated from list prices on your own key (BYOK) — not a bill.

Good as far as it goes — it is honest that the figure is an estimate. It does not say in which direction, or what is missing, and both are now known:

1. It excludes Coder Engine spend entirely (Coding Engine spend is invisible in the usage ledger — the largest real cost isn't recorded #267). The coding line reads $0.70/67 calls, which is only the cloud-side Pilot; the Engine on the user's own machine is never recorded. For a Coder user the page understates real spend substantially while looking complete.
2. Cached tokens are over-priced ~10× (Cache-read tokens are priced at the full input rate — overstates cost ~10x on the cached path #268), which pushes the other way on chat-heavy accounts.

A user reading $40.92 reasonably concludes "that is what my keys were billed". It isn't, in two directions at once, and nothing on the page says so.

Build

• A short "what this does and doesn't include" note or info popover on the Usage page:
  • platform-mediated calls on your key — chat, voice, Pilot/Co-pilot/Overseer decisions
  • not the coding CLI's own spend (that bills to your Claude Code subscription or key directly)
  • list-price estimate, not a provider bill
• Once Coding Engine spend is invisible in the usage ledger — the largest real cost isn't recorded #267 lands, replace the exclusion note with the actual engine figure and mark which engines report (Claude) vs cannot (raw Codex/Grok).
• Once Cache-read tokens are priced at the full input rate — overstates cost ~10x on the cached path #268 lands, show the cached-token share — it is the interesting number, since it is the saving.

Acceptance

• No user can read the total as a bill.
• The Engine exclusion is stated while it is true, and removed when it stops being true.

Small, and worth doing before #267/#268 rather than after — the disclosure is correct right now.

[serge-ivo]

Shipped in 32f1b37 (feat(console): disclose what the usage estimate leaves out).

A collapsed "What this includes — and what it doesn't" disclosure now sits under the existing estimate note on store/console/src/pages/Usage.tsx:

• Included — calls the platform makes on your key: chat, voice, translation, and the Pilot/Co-pilot/Overseer decisions behind a coding session.
• Not included — the coding CLI's own spend. The Engine runs on your machine and bills your Claude Code subscription or API key directly, so none of those tokens reach the page, stated with its consequence: "If you use Coder, your real spend is higher than the figure above."
• List-price estimation, with a pointer to the provider dashboard for the authoritative number.

Collapsed rather than always-open because the page's job is the numbers; the caveat has to be findable, not loud.

Verified before writing it, rather than trusting the ticket:

• The Engine exclusion is real. UsageKind in workers/api/src/lib/usage.ts has no engine member; coding is the cloud-side Pilot only. The disclosure is true as written today.
• The cache claim in this issue is wrong. CACHE_READ_MULTIPLIER = 0.1 in ai-pricing.ts already prices cache reads at a tenth of a fresh input token, so there is no ~10x over-count. Consistent with Cache-read tokens are priced at the full input rate — overstates cost ~10x on the cached path #268 being closed. No such claim was added.
• The optional "show the cached-token share" follow-up is already done — the page has shown a "Prompt cache hit" stat and a tokens served from cache at a tenth of the input price line since that stat landed. Nothing to add.

The Scope doc comment names #267 as the trigger to replace the exclusion bullet with the actual figure, so the disclosure does not outlive the gap it describes. workers/api/** and packages/browser-runner/** were left untouched — #267 is in flight in parallel.

Typecheck clean, biome identical to baseline, 3380 tests pass. No UI test added: the console has no test infra (no vitest/testing-library).