Any model can name another model as its backup. The gateway tries the
backup when the model is offline, rate limited, or its upstream fails
before the client has received anything. The response reports the
requested model, so the client sees no switch. The request log records
each failed attempt and the address that served the request, and the
cost uses the backup's prices.
The sync took meta.n_ctx_train from the model listing, which is the
training context, not the -c the server runs with. The llama.cpp provider
now asks the server for the live value: /props on a single server, the
router's /models status and /props?model=, and /upstream/<model>/props for
models llama-swap has loaded. Stopped models are never probed, because
that would start them.
Admins can also pin the context window per model from the Models page.
The override wins over the discovered value and survives syncs.
Model catalog is a table keyed by the exact provider/slug address with
a copy button. The chat model picker filters by address and name.
API Keys shows the base URL and a curl example, also right after a
key is created. /v1/models entries gain created, display_name and
context_length.
Shared tokens and components, responsive layout, real health badge,
request volume chart with hover, log detail with errors and cost,
provider enable and edit, override clearing, OIDC re-login that
returns to the current page.
extractDelta and extractCompletion read only reasoning_content, so
reasoning from OpenRouter (reasoning, reasoning_details) and Anthropic
(thinking_delta blocks) was dropped from logged completions. The
streaming path now accumulates reasoning and content separately and
wraps reasoning in think tags, matching the non-streaming format.
The dashboard Test Chat page and the Anthropic stream/translate paths
also recognize the OpenRouter and Anthropic reasoning fields.
Admin logs list no longer joins prompt/completion content for every row —
content is fetched per-row only when a log is expanded in the dashboard,
since a table of chat-length blobs was slow to serialize even at low row
counts.
request_logs.user_id (previously unpopulated despite an index existing on
it) is now set from the owning key on every insert, including synthetic
dashboard keys, and log queries filter on it directly instead of a
virtual_keys subquery. Pagination switches from OFFSET to a (created_at, id)
keyset cursor so paging deep into the log stays cheap. Adds indexes on
provider_name/key_prefix used by the admin filters.
- Migration 000003: request_log_content table (messages jsonb, completion text)
- InsertRequestLog now returns the new row's ID so content can be linked
- Chat handler captures prompt messages and completion text (streaming accumulates deltas)
- ListRequestLogs left-joins content; logs API passes key filter through
- KeyStats query aggregates requests/tokens/cost/errors per key prefix
- GET /api/admin/stats/keys endpoint
- Logs page: expandable rows show prompt + completion inline
- Analytics page: per-key usage table below the chart
Add /auth/login and /auth/callback handlers. The backend holds the
client secret and exchanges the authorization code server-side;
the browser never sees the secret. Redirect URI and post-login
fragment destination are derived solely from GATEWAY_SERVER_PUBLIC_URL
(required when OIDC is set) to prevent host-header injection.
Tailwind v4 dark design (DM Sans/Mono, violet accent). Pages: Overview,
Logs, API Keys, Providers, Models, Analytics, Routing, Settings.
Auth gate uses localStorage token validated against /api/admin/providers.
Vite builds to internal/web/dist which is embedded into the binary.
