[fish-not-phish]

What

New Cost compare page (/cost, sidebar → "Cost compare", mobile Stats sheet, command
palette): prices the fleet's token counters at any hosted model — "what would our usage have
cost on Claude Opus, Sonnet, GPT-5.2, …?"

• Any model, live: searchable dropdown over the community
  litellm price catalog
  (~2.9k chat models, prefix-ranked search). The last pick is remembered per browser.
• Honest cost split: computed prefill × input rate, completions × output rate, cached
  prefill × the provider's cache-read rate when published (else the full input rate) — so the
  page also shows what the prefix cache saves in the selected period.
• Same time ranges (24 h → all time) and the same per-Spark/port/model token counters as the
  Token totals page, so the two pages reconcile; All time uses lifetime totals.

How

• GET /api/model-prices (server/prices/ModelPrices.js): fetches the litellm JSON, keeps
  chat-shaped entries with real per-token input/output costs, serves a slim catalog. Cached in
  memory for 24 h; the last good copy persists to config/model-prices.json (gitignored) and
  is served if the upstream fetch fails — stale prices beat an error, 502 only with no cache
  at all.
• All pricing math lives in src/components/CostPage/costStats.ts — pure functions, no
  React, no I/O — so the page component stays presentation-only.

Testing

Per README.md §Contributions:

• npm run typecheck — clean
• npm run test:frontend — 99 files / 803 tests pass, including the new
  src/components/CostPage/costStats.test.ts and the updated MobileTabBar sheet test
• Server: node --test server/__tests__/isolatedEnv.test.js server/prices/__tests__/ModelPrices.test.js — parser, disk-cache fallback and the
  no-cache 502 path covered; server/prices added to the test:server glob in
  package.json; the new MODEL_PRICES_JSON_PATH registered in isolatedEnv.js
• Full server suite run on the branch: no new failures vs main (the deltas are the
  environment-dependent sudo/vulnhunt/llmlaunch checks, identical on pristine main)
• Manual: served endpoint returns 2865 models; /cost renders in all four themes; the
  page keeps working with the network path down via the disk cache

Notes

• No new runtime dependencies; the fetch uses global fetch (Node 18+) like
  server/collectors/HermesReleases.js.
• The catalog refresh is lazy (on request, TTL-gated) — no timer, nothing to start/stop;
  server/index.js changes are two lines (import + one app.get).
• The shared Spark model is untouched; costs aggregate the existing fleet-wide token rows.

[MiaAI-Lab]

Reviewed at ff4cd86. It merges cleanly and typecheck, build and the full tests pass (server 1064/1065, frontend 803), and I found no injection or SSRF problem (the URL is a constant and the route sits behind the auth middleware), but it is not ported yet. (1) server/prices/ModelPrices.js:124-129 seeds the memory copy from the disk cache, so after the first failed refresh the next request serves that old catalog with stale:false for the whole TTL, hiding that prices are out of date; keep the disk copy separate and answer stale:true while refreshes fail. (2) A failed refresh is retried on every request with a 15 s timeout and no shared in-flight request, and CostPage.tsx:54-60 loads prices and usage together, so a LAN without internet leaves the page on 'Loading prices…'; add a short negative cache, one in-flight request, a shorter timeout and load prices independently of usage. (3) CostPage.tsx:293-294 divides an All-time cost by a fixed 30 days; reuse the tracked-days helper from Token totals. (4) The page calls raw.githubusercontent.com with no way to turn it off; please make it opt-in (setting or env), document it in the README, and add redirect: "error" and a body size cap. Leaving it open for the author.

(Review notes prepared with Claude Code on behalf of the maintainer.)

[fish-not-phish]

Thanks for the careful review! All four points are addressed in d95f735:

1. Stale flag — the disk fallback is no longer promoted into the memory cache. It's a separate copy that answers stale: true for as long as refreshes keep failing; only a successful fetch resets to stale: false.
2. Refresh resilience — failed refreshes are negative-cached for 60 s and share a single in-flight request, and the timeout drops to 5 s. The page also loads prices independently of usage now, so a network-less LAN gets the usage view plus a "price catalog could not be loaded — Retry" note instead of a spinner.
3. All-time average — now divides by trackedBuckets(firstKey, "day", …), the same rule the Token totals page uses, instead of a fixed 30.
4. Opt-in fetch — the GitHub refresh only runs with MODEL_PRICES_FETCH=1 (documented in README Environment variables and .env.example); while off, the route serves the disk cache and makes no requests. The fetch also uses redirect: "error" and enforces a 10 MB body cap (content-length and actual bytes).

Tests: five new cases in server/prices/__tests__/ModelPrices.test.js covering each of the above (stale persistence, retry window + in-flight dedupe, opt-out 502 hint, redirect/cap); npm run typecheck clean; frontend suite 802 passing — the one failure is a pre-existing timezone-dependent flake in benchShareCard that fails on pristine main too.

[MiaAI-Lab]

Now on main as b361e06, squashed with your authorship (thank you). I re-checked the four points from the review at server/prices/ModelPrices.js and CostPage.tsx: the disk copy no longer masks staleness, failed refreshes are negative-cached with one request in flight, the All-time average uses the tracked-days helper, and the GitHub fetch is opt-in (MODEL_PRICES_FETCH=1, server/config.js:44) with redirect: "error". One change on top: the 10 MB cap was checked after res.text() had buffered the whole body, so a response without Content-Length (or a compressed one) got past it. fetchJsonCapped (ModelPrices.js:77) now counts bytes as they stream and cancels past the cap, with a test. I also added the CHANGELOG entry. Typecheck, build and the full tests pass.

(Review notes prepared with Claude Code on behalf of the maintainer.)