Both surfaces wrote zero bytes for the whole time-to-first-token, so an
intermediary with an idle read timeout drops a healthy connection before the
first token. They reached neither keepalive engine, which is what #37322 left
open.
The streaming assistants run spends that wait inside the awaited call that
produces its response, since create_response buffers the first chunk, so it
takes the same open_sse_before_first_byte seam the native routes use. The A2A
route only contacts the upstream agent once its body iterator is first pulled,
so nothing is awaited before the response exists and the gap has to be filled
from inside the stream instead; wrap_sse_stream_with_keepalive_pings already
does that and now takes the filler as a parameter, so A2A gets an SSE comment
its JSON-RPC clients discard rather than Anthropic's ping event.
Off until an operator sets litellm_settings.sse_keepalive_ping_interval_seconds.
* fix(otel): bound and shut down credential-scoped tracer providers
Each credential-scoped TracerProvider owns a BatchSpanProcessor worker thread that
only stops on shutdown, and the v1 cache holding them was an unbounded, unsynchronized
dict that never shut anything down. Every distinct team/key credential set therefore
added a thread for the life of the process, and concurrent first-requests for the same
credential set orphaned duplicate providers outright.
Make the cache a lock-guarded bounded LRU that shuts down whatever it drops, matching
the v2 TenantTracerCache. Providers wrapping a caller-supplied SpanExporter instance
share that exporter with the logger's own provider, so they are dropped without
shutdown; those use SimpleSpanProcessor and own no thread.
* fix(otel): reclaim dropped providers on a dedicated executor
Sustained credential churn queues one blocking shutdown per eviction, so using the
shared logging executor let an unreachable tenant endpoint stall unrelated logging
work behind the OTLP retry budget. Give provider shutdown its own bounded pool; its
threads spawn lazily, so a proxy that never evicts still pays nothing.
* fix(otel): decide provider shutdown from the victim, not the evicting request
Both dynamic entry points share one provider cache, so it can hold providers of
mixed exporter ownership. Reading the ownership flag from the evicting request
therefore stopped a shared caller-supplied exporter in one direction, silencing
telemetry process-wide, and leaked a BatchSpanProcessor thread in the other.
Cache ownership alongside the provider so the drop decision reads the victim's
own flag.
* fix(otel): honor the widened header mapping type instead of dict only
Widening the header parameter to Mapping left the isinstance check on dict, so a
non-dict Mapping silently returned no headers at all, which for the OTLP path means
an unauthenticated exporter and no traces with nothing raised. The dict branch also
returned the caller's own object, and dropping the defensive copy at the call site
let that alias reach a long-lived exporter. Match on Mapping and copy.
* fix(otel): do not give a provider we may never stop an interpreter-exit hook
Every TracerProvider registers an atexit hook by default, and that hook holds a strong
reference. Providers wrapping a caller-supplied exporter are dropped without shutdown,
so they stayed pinned for the life of the process and then stopped the shared exporter
at exit. Tie shutdown_on_exit to ownership: those providers use SimpleSpanProcessor and
buffer nothing, so they lose no flush, while providers that own their exporter keep the
hook and their exit flush.
Also stop the victim the eviction test leaves behind, and trim the added comments.
Key and team router_settings set enable_tag_filtering on the request kwargs,
and get_deployments_for_tag already treats that as authoritative, but
_select_pre_routing_strategy only consulted the router-wide flag, so tagged
auto-router markers still captured untagged requests from keys that enabled
filtering. The e2e auto-router module now enables tag filtering through
key-level router_settings instead of flipping /config/update module-wide,
which was denying concurrently running tagged requests from other suites on
the shared per-build CI proxy.
* refactor(ui): move the model info view and pass-through endpoint forms off tremor
The pass-through settings form's Save button relied on tremor's implicit
submit, so it now carries an explicit type="submit" because the Base UI
button defaults to type="button". The include-subpath switch inside the
antd Form.Item is wired through onCheckedChange plus form.setFieldsValue,
since the Base UI switch does not read the onChange antd injects. Both
tab strips keep every panel mounted so the edit forms survive a tab
switch, pinned by a new mount-contract test on the model info view. Also
prunes the six tremor no-restricted-imports suppressions these files no
longer need.
* test(ui): pin the model info overview panel to its own dom node across tab switches
* fix(ui): keep the line tab strip on the model info and pass-through views
Both tab strips were bare tremor TabLists, which defaulted to the line
variant, so the straight rename turned them into filled segmented pills.
They now use the same full-width line strip the agent, guardrail and
prompt info views ship.
The antd form reset on [form, actualSchema, tool], so it reseeded whenever
the schema changed and not only when a different tool was picked. Keying the
migrated form's remount on tool.name alone narrowed that: a same-named tool
whose schema changed would have kept its old indexed values and submitted
them under the new schema's keys, and any field the new schema added would
never get its default. The key now covers the schema content as well.
Reachability, so nobody reads more into this than is there: selectedTool is a
state snapshot set on click, so this is not reachable through the current
parent. It is a latent divergence rather than a live bug, and it is fixed
because the contract for this migration is zero functional change.
The schema default helpers move into the pure module beside the other
argument logic. They arrived carrying the original's explicit any and built
their result by mutating it; they are now typed with unknown, built by
spread, and covered directly by unit tests rather than only through a render.
Also takes the last five hardcoded neutrals onto tokens. The earlier pass
scanned the gray family only and did not see the slate ones in the tool name
chip, which had no dark variant. Counting colour utilities that have no
dark: counterpart reads 0 for this file now, against 64 before the migration.
The Base UI combobox only renders a selected value that is present in its
item list, so the port synthesizes an item for the current value when the
search results no longer contain it. That synthetic item carried an empty
user, and selecting it ran the same handler as a real result, wiping both
the email and the user id before submit.
The synthetic item now carries no user at all and the select handler
ignores it, so reselecting the value already in the field leaves both
identity fields alone. antd needed none of this: its Select renders a
value that is absent from its options.
The team create and edit forms send a different set of keys depending on
which collapsible sections the user opened, because a closed section is
unmounted and its values never reach the request. Nothing covered that,
so a form rewrite could change the request body without failing a test.
Pins the exact key set the create form sends with every section closed,
the keys Additional Settings adds once opened, and that a value typed
then re-hidden is dropped while a reopened one is restored. Does the same
for the team member and search tool sections on the edit form, asserting
absence at the wire level rather than just comparing values.
Also hardens two option queries in the member modal suite onto the option
role, and lifts the duplicated mock seeding in the team info suite into
one function both blocks call.
* test(ui): raise vitest test and hook timeouts for CI headroom
The UI unit suite runs about 3x slower on the CI runner than locally, which
put the slowest cases right on the 30s per-test limit. TeamInfo's pass
through routes case takes ~8s locally and has been failing on staging at the
timeout across consecutive runs even though it passes reliably when run
directly.
Raise testTimeout to 60s and set hookTimeout to 30s so the current slow cases
have headroom. This is a stopgap while the suite gets split into proper tiers,
not a fix for the underlying per-test cost.
* test(ui): query agent form panels with findByRole like the rest of the file
The panel helper was the only synchronous query in add_agent_form's
integration test; every other lookup already retries via findBy. On the CI
runner the second case has been failing with "Unable to find an accessible
element with the role button and name /Cost Configuration/" against a modal
whose body had not rendered.
Make the helper retry like its siblings and await it at each call site.
The tool test panel drove its argument fields through an antd Form, so the
call payload was whatever rc-field-form happened to have mounted. It now runs
on react-hook-form with shadcn controls, and the payload itself lives in
toolCallArguments.ts as a pure function of the schema fields plus the entered
values.
Fields bind by index rather than by name, because an MCP tool's JSON schema
can name a property anything: a key containing a dot would be one flat key to
antd but a nested path to react-hook-form. Binding to args.0, args.1 and
zipping back to the real keys at submit time keeps the emitted arguments
identical whatever the server calls its properties.
Coercion, the blank filter, the required and JSON rules, and the params
wrapper for nested-object schemas all keep their previous behaviour, and the
neutral colours in the panel move onto tokens so it reads correctly in dark
mode.
* refactor(ui): move the admin, SSO, SCIM, alerting and fallback forms off tremor
Swaps the tremor Button, Card, Callout, Grid, Divider, Text, Title, TextInput, Table parts, Badge, Icon and Switch in these nine files for the shadcn layer and lucide icons, keeping antd in place. The SCIM create-token button keeps an explicit type="submit"; the two SCIM copy buttons sit outside the antd form so they stay plain buttons, and the alerting form's Enterprise Feature upsell button is a link wrapper rather than a save action, so it deliberately stays type="button" while the form keeps its own Update Settings submit. The teal login callout on the admin panel maps to the info Alert variant since no teal variant exists. Prunes the nine tremor no-restricted-imports suppressions these files no longer need.
* fix(ui): keep the alerting settings name column left aligned
tremor's TableCell hardcoded text-left, so the align="center" attribute never took effect. The shadcn cell has no text-align of its own, so translating that attribute into text-center would have centered the field name and its description for the first time.
* fix(ui): restore the CloudZero key reveal toggle and the SCIM divider gap
tremor's TextInput drew its own show/hide button whenever the type was
password and the field was not disabled, so the straight pass-through to
the plain shadcn Input silently deleted that affordance from the CloudZero
API key field. Rebuilds it with the InputGroup reveal pattern that
email_settings.tsx already uses, behind a small local control so the antd
Form.Item keeps its id, value and onChange wiring and the field still
matches its shadcn sibling in the same form.
Also puts the SCIM separator back on my-6: tremor's Divider was
"w-full mx-auto my-6", and the conversion shipped my-4, tightening that
gap by 8px on each side. The disabled SCIM token field stays a bare Input
because tremor suppressed its toggle when disabled too.
The selector takes value and onChange, which is exactly what lets it sit
inside both an antd Form.Item and a react-hook-form FormField, so the doc
bullet naming only antd was about to describe half the truth. The props
interface already states the contract.
That bullet was also the only Form.Item match in the file, and it has twice
inflated the migration's canonical tag count, which now needs no subtraction.
Its four hardcoded colours move to tokens for the same reason as the rest of
the sweep. Its antd Select stays.
The agent forms moved to react-hook-form in #37357, which deliberately kept
the antd Select for the agent type picker because its dropdownRender footer
and two-tier options have no shadcn equivalent. The Form.Item wrapped around
it survived as a side effect rather than for that reason: it carries no name
and no rules, and the file renders no antd Form at all, so it bound nothing
and validated nothing.
It was also the only label in the file rendering antd's required asterisk,
while AgentFormField renders genuinely required fields without one. Moving it
to the file's own Field, FieldLabel and labelWithHint makes it match, and
gives the label a control to point at.
The shared alias editor was light-only: 18 hardcoded palette classes across
its headings, table cells, config preview and five raw controls, so it read
grey-on-grey against the dark dashboard theme.
Its five buttons also carried no type attribute, and all three consumers
render it inside a real antd Form with an onFinish. Measured against the
pre-change file in jsdom with a liveness gate on both sides: clicking Add
Alias fired the parent's onFinish once, and a row action fired it once more.
So editing a model alias inside the create-team, team-update or create-key
form also submitted that form. shadcn Button renders type="button", which
closes the path.
Swapping the raw input and button elements for the shadcn primitives is what
makes the colours resolve, since a bare element needs the whole token set
hand-written to work in both themes.
`items` feeds `hasNullItemLabel` as well as label resolution, and that
selector decides whether `Select.Value` renders the placeholder. None of
the twenty fixed sites has a null-valued item, so their placeholders are
unchanged, but nothing pinned that.
Cover both halves: the placeholder still renders when nothing is
selected, and a null-valued entry in `items` takes over from it. The
second case is the trap to avoid when adding `items` to a Select that
relies on its placeholder.
Base UI's Select.Value resolves an option's label only when the root
carries an `items` prop or the Value has a child. `resolveSelectedLabel`
in @base-ui/react/internals/resolveValueLabel.js falls through every
branch to `stringifyAsLabel(value)` otherwise, and `state.items` is
written only from the root's `items` prop, so the `<SelectItem>` children
rendered inside `<SelectContent>` never populate it.
A self-closing `<SelectValue />` on a root without `items` therefore
renders the raw value once something is selected. The placeholder branch
still works, so the trigger looked right until the user picked an option
and then showed `development` for Development, `LiteLLM_VerificationToken`
for Keys, `all` for All Actions, and `24h` for Daily.
Pass `items` at the 20 affected sites, using the array form the other 52
call sites already use. Where a literal option sat alongside mapped ones,
build one array and map the options over it so the labels and `items`
cannot drift.
The record-map form is avoided deliberately: `items[value]` on an object
literal reaches Object.prototype, so a dynamic value named `toString`
would resolve to a function and React would throw on it. The array form
matches with `.find` and has no prototype lookup, which matters where the
values are user-supplied model groups, team ids and key aliases.
Also replace the option lookup in CompetitorIntentConfiguration's test
helper, which searched by text and clicked the last match. That match is
now ambiguous because the trigger carries the label too, and the helper
already flaked roughly one run in six before this change.
Ports user_search_modal from antd Form to react-hook-form plus the shadcn
kit, keeping the antd Modal and Alert shells. Payload parity was proven by
rendering the antd original beside the migration in one describe.each: an
untouched submit yields the same three keys with the identity fields
undefined, and picking an option yields the same email and id on both sides.
antd Select swallows Enter, so the original never submitted from a field.
The Base UI combobox does not, which added an Enter-to-submit path; the
inputs now swallow Enter and both sides measure zero submits from every
field with one from the button.
* refactor(ui): move the budget, cache, cost tracking and playground forms off tremor
Swaps tremor Accordion for the Base UI Collapsible, TextInput for the shadcn Input and the two
tremor Buttons for the shadcn Button across the budget modals, cache settings, the cost tracking
add-provider and add-margin forms and the playground model selector. The accordion bodies keep
tremor's unmount-when-closed semantics, since headless-ui's Disclosure.Panel and Base UI's panel
both default to unmounting, so the antd fields inside behave exactly as before.
The two cost tracking buttons are the one deliberate behaviour change. tremor's Button renders a
bare button with no type, so inside the antd Form that wraps both components it was an implicit
submit on top of its own onClick. For the discount form that meant every click ran
handleAddProvider twice, once from onClick and once from the form's onFinish, and for the margin
form the submit did nothing at all because that Form has no onFinish. The shadcn Button forces
type="button", so the add now fires once from onClick alone and no type="submit" is added back.
The three inputs that used onValueChange now read e.target.value, and each one gained a test that
types into it and asserts the reported string, so the wiring cannot silently regress. The cache
settings suite gained a collapse contract test that the advanced sections are absent until the
section is expanded. Prunes the six no-restricted-imports suppressions these files no longer need,
each dropping from two to one for the antd import that stays.
* fix(ui): keep enter to submit on the cost tracking add forms
The shadcn Button forces type="button", so converting the two tremor buttons left both cost
tracking modals with no submit button at all. Each form still holds two fields that block
implicit submission, the provider select's search input and the value input, so pressing Enter
stopped adding anything. Both buttons get type="submit" back.
For the discount modal that alone would restore the double add the conversion had just removed,
since a submit also ran the form's onFinish, so the parent drops onFinish and the now dead
handleFormSubmit. Click and Enter both go through onClick exactly once. The margin form's parent
never had an onFinish, so restoring the submit type there is enough on its own.
Adds three cases to the cost tracking settings suite: the discount add fires once from a click,
the discount add fires once from Enter, and the margin add fires once from Enter. Dropping either
type="submit" kills the Enter cases and putting onFinish back makes both discount cases see two
calls. Also drops the two empty placeholders on the budget modals that only existed to suppress
tremor's "Type..." default.
* fix(ui): restore Enter-to-submit on the margin modal
The tremor Button rendered a bare native button, which defaults to
type="submit", so Enter in the percentage field submitted the margin
modal. The shadcn Button wraps Base UI, which defaults to type="button",
and the migration also replaced the margin modal's form element with a
plain div, so Enter went inert while the visually identical discount
modal kept working.
Give the margin modal the same form wrapper the discount modal already
has and mark its action button as the submit button. Also move the cache
settings advanced-section test into the integration file, where a test
that renders the real component tree belongs.
* refactor(ui): move the teams page and team detail views off tremor
Swaps the tremor Accordion, Badge, Button, Card, Grid, Text, TextInput and
Title usages in Teams.tsx, TeamInfo.tsx, EditMembership.tsx and
LoggingSettings.tsx for the shadcn layer. The team model badge colour map
becomes a variant map: all-proxy, direct and access-group chips render as
secondary and no-default as outline, so the kind is now conveyed by the
tooltip rather than by hue. The LoggingSettings top decoration is drawn with
border-t-4 border-t-blue-500 and its light red Remove button becomes a ghost
button with red text. antd stays in place for this pass and the eslint
no-restricted-imports counts for the four files ratchet down by one each.
* fix(ui): keep the password reveal and model badge hues in the team views
The tremor TextInput rendered a show/hide button for every password field, so
the shadcn swap silently dropped it for the sensitive logging parameters. The
password branch now renders an InputGroup with an eye toggle, matching the
pattern email settings already uses, and a test pins the masking.
The team model chips go back to four distinct colours by way of the shared
StatusBadge, so a directly granted model still reads differently from an
access group one without hovering for the tooltip.
The hand-drawn blue accent on the logging integration card is dropped: the
tremor decoration it replaced never rendered, because the caller's own border
classes won the class merge, so the bar was new rather than preserved.
* fix(ui): drop the dead empty placeholder on the team name field
The team name input carried placeholder="" only to suppress tremor
TextInput's default "Type..." hint. shadcn Input has no default
placeholder, so the empty string does nothing and the field now relies on
its label, matching the other converted create-team fields.
A terminal HTTP failure partway through chunking now logs the summed usage
and cost of the ApplyGuardrail calls AWS already billed, mirroring the
blocked-chunk path.
Removes one named callback from a team and leaves the team's other callbacks
registered and firing. Before this, the only removal route was
POST /team/{team_id}/disable_logging, which clears every callback at once, so a
tenant sharing a team could not deregister its own integration
The handler filters metadata["logging"], keeps the survivors encrypted, refreshes
the cached team so the removal applies to keys that are already live, and emits a
redacted audit row, matching what the add and disable routes do
Resolves LIT-5161
A blocked chunk now logs the summed usage and cost of every ApplyGuardrail
call AWS billed for the logical request, not just the blocking chunk.
Client-supplied metadata.standard_logging_guardrail_information is stripped
at the proxy boundary so callers cannot forge (even negative) guardrail
cost into spend, and guardrail_information_cost ignores negative or
non-finite entry costs as defense in depth. The cost map schema test now
allows guardrail_cost_per_unit and the guardrail mode.
