[Console v2] /console/api/request-logs: 503 ServiceUnavailable above ~20-50 page size, /request-logs/export always 503 — and the legacy per-key /usage/rows endpoint is gone #51806

Summary

Two regressions in the Console v2 usage/logs API, plus one retired endpoint that third-party usage dashboards depended on:

1. GET /console/api/request-logs returns 503 ServiceUnavailable as soon as the page size goes above ~20–50 (the threshold drifts during the day).
2. GET /console/api/request-logs/export returns 503 ServiceUnavailable every time.
3. The legacy GET /console/api/usage/rows (which returned one row per request with costMicroCents + model + serviceApiKeyId) now 404s for every query, which removed the only way to break usage down per API key.

Reproduced from a normal user session (Go subscription, signed-in Console cookies), no VPN, macOS.

Environment

• Client: custom usage dashboard; auth via Console cookies (auth, __Host-console_session) + x-org-id header (same as the Console web app).
• Region: China Mobile (CN) — but see "healthy endpoints" below, this does not look like a network issue.
• Date: 2026-09-28, 16:33 UTC+8.

Repro A — page size threshold on /usage/request-logs

Same time window for every call (today 00:00 → +24h, UTC+8):

curl 'https://opencode.ai/console/api/request-logs?limit=100&since=<ms>&until=<ms>' \
  -H 'Cookie: auth=…; __Host-console_session=…' -H 'x-org-id: <redacted>'
# → HTTP 503 {"_tag":"ServiceUnavailable"}

Note the threshold is not stable: earlier the same day limit=20 also returned 503 while limit=10 worked; at 16:33 limit=20 worked and limit=50 did not. Looks like a server-side cost/timeout guard rather than a hard limit.

Repro B — export endpoint

curl 'https://opencode.ai/console/api/request-logs/export?format=json&since=<ms>&until=<ms>' -H 'Cookie: …' -H 'x-org-id: <redacted>'
# → HTTP 503 {"_tag":"ServiceUnavailable"}   (every attempt, incl. a single day)

Repro C — legacy per-key endpoint is gone

curl 'https://opencode.ai/console/api/usage/rows?range=30d&pageSize=100' -H 'Cookie: …' -H 'x-org-id: <redacted>'  # → 404
curl 'https://opencode.ai/console/api/usage/rows?range=7d'                                                          # → 404

That endpoint used to be the only one carrying serviceApiKeyId per request, i.e. the only way to show cost/tokens per API key. request-logs carries serviceAPIKeyID too, but it is (a) currently unusable beyond ~20 rows per page, and (b) has no data older than ~Sept 19 for this workspace, while usage/cost-by-day has 30+ days.

Healthy endpoints (same minute, same cookies)

usage/summary, usage/cost-by-day (day + bucket=hour), usage/models (incl. since=…), usage/users, service-accounts — all 200 in 0.8–3.6 s. So the problem is isolated to the two logs endpoints.

Questions / asks

1. Can request-logs accept normal page sizes (100, as used by the Console's own Logs page) again — or is there a documented max?
2. Can request-logs/export be fixed? It is the only practical way to pull a day/batch without hundreds of small pages.
3. Is the legacy /usage/rows coming back? If not, could the per-key dimension be exposed somewhere cheaper, e.g. usage/models?groupBy=service_api_key, a serviceAPIKeyID column in usage/cost-by-day, or a scope in v2/usage/export? Right now there is no endpoint that returns per-key totals for a 30-day window.

Related reports (other projects adapting to the Console v2 migration)

• [BUG] Go quota sidebar fails after Console migration Wikijito7/opencode-model-usage#48 — "Go quota sidebar fails after Console migration"
• feat: restore OpenCode Zen Usage Snapshot on v2 martinvidovic/opencode-limits#48 — "restore OpenCode Zen Usage Snapshot on v2"
• [bug]: budgets/org returns null without an org budget → permanent "Could not parse" error for opencode-zen slkiser/opencode-quota#297 — budgets/org shape change

(Account identifiers intentionally redacted; happy to provide workspace/service-account IDs privately if a maintainer needs them.)