feat(models): labelled capability estimates, automatic prices with their source, and a Usage page that loads - #8738

Summary

Models reached through relays stop reading as "unknown" everywhere, and costs stop reading as "unknown" for most traffic. For a model whose context window, image input or effort levels are not reported, the dashboard now shows a labelled estimate from the same model id elsewhere, with where it came from, and a Test it probe can check what the relay actually accepts. Prices are fetched automatically from models.dev and OpenRouter, every price shows its source, and a manually set price always wins. The Usage & Logs page loads in seconds instead of timing out on large logs, its models table fits the window without horizontal scrolling, and apps that used to be lumped into "Unclassified" and "Unknown" get names.

This PR gathers three earlier drafts about model metadata and usage so they can be reviewed and merged as one change on current dev, with no stacking: #8674, #8708 and #8728 (now closed; their descriptions keep the full design notes). There is one commit per original PR, then the follow-ups listed below.

Sponsored surfaces. This PR needs maintainer-sponsored for src/server/auth-cors.ts only: two config-key role entries (modelCapabilityProbes: "editor", priceMultiplier: "editor") and the priceMultiplier range check in the config validator.

Parts

1. Capability inference, provenance and probe (was #8674).

• src/providers/model-capability-inference.ts looks a model id up per axis (context window, input modalities, reasoning efforts) in order of authority: original-vendor registry entries (multi-vendor hosts excluded), bundled vendor metadata, other providers' explicit declarations, bundled aggregator metadata. Ids match ignoring case, a vendor/ prefix, dot-versus-hyphen and a [...] suffix; disagreeing sources of one authority give no answer.

• GET /api/models routed rows gain capabilitySources (manual > probed > reported > inferred > unknown) and inferredCapabilities with their basis. Estimates are display data: they never route and never enter the Codex catalog on their own (the Models GUI cannot define a fallback context window for metadata-poor third-party providers #1073 decision stands); applying one means saving it as a declaration. No second window override is added (modelContextWindows already is one).

• POST /api/providers/probe-capabilities reuses the completion diagnostic (32-token cap, 20 s, no retries): a plain baseline first, then a 1x1 PNG and one request per effort. Metered destinations send nothing without confirmMetered: true and return plannedRequests. With apply: true conclusive results are written and labelled in a sparse modelCapabilityProbes; a later manual edit drops the label. Only status codes reach the response; upstream bodies are never returned or logged.

2. Automatic prices with their source (was #8708).

• Resolution order: manual modelCosts > official > models.dev (original-vendor entries only) > OpenRouter (per-token converted to per-1M) > unknown. Catalog prices are verified-derived, so estimates using them are marked estimated and carry the feed URL and fetch time. A feed with two prices for one key drops the key.

• Cached in <configDir>/price-catalog-cache.json; a daily refresh is triggered by dashboard reads, and a pricing miss queues a background fetch with a 6 h cooldown per missing key. Nothing runs on or awaits the network on the request path. Fetches are anonymous GETs with redirect: "error", a 15 s timeout and a 32 MiB cap.

• providers.<name>.priceMultiplier (opt-in, > 0 and <= 1000) scales automatic prices for relays that charge a multiple of list price; never manual ones. OAuth/forward providers and Codex accounts carry apiEquivalent.

• /api/models rows gain price; new POST /api/pricing/refresh and PUT /api/providers/:name/price-multiplier. The dashboard shows a price-source chip on every model row and the automatic price in the price dialog with Adopt as manual and Refresh prices.

• Also registers POST /api/providers/probe-capabilities in route-registry.ts (part 1 added the route).

3. Usage page speed, table fit and app names (was #8728).

• The dashboard always sends client=all&device=all. The route treated that as a filtered query, bypassed the warmed cache and rebuilt a large SQLite projection. It now serves that selection from the shared unfiltered aggregate (src/server/management/usage-source-presentation.ts). On a synthetic 270k-row / 745 MB ledger the cold request went from 23.8–64.5 s (past the page's 30 s timeout) to 12.4 s at start and 5.1 s once startup settles; cached it is 0.01 s.

• The models table is seven columns, each secondary figure on a second line; only the model column stays pinned.

• Rows written before client tracking are split by the API they arrived on. Requests from unrecognized clients record a bounded product name (first product token, lowercased, versionless, ^[a-z0-9][a-z0-9._-]{0,39}$; never the header value), at most 24 per report.

• The "N requests excluded" caption and the coverage panel explain why requests are missing a cost.

Follow-ups in this PR.

• Design review fixes (3edcea648): price-source chips now mark only the exceptions (manual, models.dev, OpenRouter, unknown) and only on routed rows, instead of putting "Official price" on every row including native Codex ones; the official rate is still in the price dialog. Rates and fetch times format in the dashboard's language rather than the browser's. The multiplier help said it scales only official prices while the server scales every automatic price; corrected and shortened in all eleven locales. The "N requests excluded" hover now states each reason only when its count is above zero and points to the Models page, where prices are set. The "before client tracking / unknown" note shows only when those rows exist; coverage category explanations are readable by screen readers, not only on hover; .usage-cell-sub is one rule instead of two conflicting ones.

• Two small additions on the same theme (ede2304e5, d9b11ba47): requests with an antigravity/hub/… User-Agent are classified as the app "Antigravity" and google-genai-sdk/… as "Antigravity CLI (agy)" (the bounded USAGE_CLIENTS roster and its GUI consumers; product names are not translated, like Codex and Grok); and verified prices for cursor/composer-2.5 (0.5 / 2.5, cache read 0.2), cursor/composer-2.5-fast (3 / 15, cache read 0.5), both from cursor.com/docs/models, plus workbuddy/cn:kimi-k3-1 mapped to Moonshot's K3.1 Pro list price (2 / 10, cache read 0.2) as verified-derived because the mapping is inferred. openai/codex-auto-review stays unpriced: OpenAI publishes no model name or price for it.

Screenshots

Local test instance with fake providers and fake data.

Verification

Branch built from dev 873ec9b by squash-merging each original PR in order; the resulting tree was compared byte for byte with a real merge of dev and the original branch tips before the follow-up commits. All runs below on this head, Windows 11, with OCX_GIT_CONFIG_GUARD_CHECKOUT pointing at a directory without .git (this machine's git 2.56 rejects GIT_CONFIG_GLOBAL=NUL).

• bun run typecheck, bun run structure:check, bun run privacy:scan, bun run skill:surface:check: exit 0.

• GUI: bunx tsc -b, bun run lint, bun run lint:i18n, bun run build:gui: exit 0.

• GUI tests (models-price-editor, usage-app-breakdown, usage-layout, usage-custom-range, usage-source-picker): 236 pass, 0 fail.

• Server tests for the theme (capability inference, probe and probe sources; the three price-catalog files; price-catalog-management-api; usage-client-attribution): 153 pass, 0 fail. tests/server/api-usage.test.ts: 35 pass, 0 fail.

• After the two small additions: usage-client-attribution, price-catalog-resolution, usage-filter-index, ci-workflows/file-size-ratchet, test-layout: 114 pass, 0 fail.

• Usage report timing (from fix(usage): load the Usage report from the shared aggregate, fit the models table, name apps #8728): synthetic 270k-row / 745 MB ledger, cold report request 23.8–64.5 s before, 12.4 s right after start and 5.1 s once startup settles after; cached 0.01 s.

• Full-suite exception: bun run test and the whole GUI suite were not run locally; several worktrees share this machine and full runs time out in unrelated modules under that load. Coverage beyond the files above is left to CI. The usage routes are reached through subprocess-started servers in some suites, so ci:full is appropriate.

CodeRabbit round 1 (18 findings: 17 inline + 1 outside the diff; all fixed, 0 declined; commits 618987db2..7d5d5a57a):

• App names: a caller that sends its own originator with the generic google-genai-sdk/ User-Agent keeps its own name (a bare SDK caller still counts as agy); a named app the full partition already holds keeps counting under its name; addRequestLog drops an invalid clientName so /api/logs and usage.jsonl agree.
• Pricing: pooled account labels get their provider's multiplier; a memoized miss is retried after the 6-hour cooldown; a failed source does not start a cooldown; a miss that joins another's in-flight fetch gets its own model merged; an explicit multiplier write (including clearing it) wins over a concurrent edit on disk; account-bound native rows are priced as their underlying model. Price dialog: a list-refresh retry after "Refresh prices" returns to editing and keeps the rate draft; a multiplier save with no response goes through the existing "outcome unknown" read-back.
• Capability probe: results are written under the native model id, a textless baseline still counts as accepted, and a change of adapter, base URL, auth mode or alias target during the probe writes nothing and answers 409. Probe labels: dropped when load repairs the axis or a save changes the declaration; a same-destination provider save keeps a probe-written reasoning ladder and its label (a hand-set ladder is still dropped by a provider-form save, as before this PR).
• Guard (016d36284): tests/usage/usage-client-attribution.test.ts fails on any client: classifyUsageClient(...) request-log context, so an inbound route that records only the closed label cannot land. feat(integrations): Gemini API inbound and one-click Antigravity app and agy CLI connection #8742 adds a Gemini route that does exactly that because the name field does not exist on dev yet; whichever of the two lands second switches it to ...usageClientLogContext(req.headers).
• Re-run on this head: usage-client-attribution + ci-workflows/file-size-ratchet 35 pass; price-catalog-miss 14, price-catalog-resolution 18, price-catalog 24, server/price-catalog-management-api 19, codex-integration/model-visibility-management-api 25, provider-capability-probe 41, provider-capability-probe-provenance 18, test-layout 2 pass; GUI models-price-editor 31 pass (the two new cases fail without the fix); bun run typecheck, structure:check, GUI tsc -b, lint, lint:i18n: exit 0.

Checklist

• Scope stays focused and avoids unrelated cleanup.

• Docs or release notes were updated when needed.

• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

Review readiness checklist

This PR stays in draft until every box below is ticked. Tick all four boxes once the requirements are met:

• Required local validation passed; commands, results, and any full-suite exception are documented.

• I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).

• I resolved all correct Codex and CodeRabbit findings.

• My PR is ready for review.

Summary by CodeRabbit

• New Features
  • Model prices can now use verified official rates or estimates from public pricing catalogs. Provider-specific multipliers can adjust automatic estimates, and eligible prices can be saved as manual overrides.
  • Model listings show price sources and capability details, including inferred values. Provider capability checks can test image and reasoning support, with an option to save conclusive results.
  • Usage reports recognize Antigravity and Google GenAI clients, distinguish named and legacy app entries, and provide clearer cost and coverage details.
• Documentation
  • Updated dashboard, configuration, and API references with pricing, capability, and usage information.