* feat(complexity_router): escalate oversized prompts to a tier that fits before dispatch
The classifier scores complexity and never prompt size, so a long agentic
session whose newest ask is trivial classifies SIMPLE onto a small-window
tier and the provider rejects it with a context-window 400 that nothing
retries. The gate runs after classification on every decision path
(classify tail and session-affinity pin), estimates prompt tokens
including the out-of-band carriers (top-level system, tools,
instructions), and when the decided tier provably cannot hold the prompt
moves the request to the lowest configured tier with a model whose
declared window fits, restricting the pick to fitting models when the
decided tier can keep it. Models with no resolvable window are never
escalated away from or onto, escalated decisions are never written as
session pins, and the decision records context_escalated plus the
original tier in spend logs.
Resolves LIT-6503
* fix(complexity_router): judge groups by smallest window, bound skips by bytes, filter adaptive picks
Review-round rework, one mechanism per finding. A group is judged by its
smallest resolvable deployment window, since the core router picks within
a group with no fit check. The counting skip is gated on UTF-8 byte
length, which BPE token counts can never exceed, so token-dense scripts
cannot slip past it; only a real tokenizer count ever moves a request and
a failed count leaves the placement alone. The fit facts now filter every
adaptive phase including cold start and the tier fallbacks. Window
questions adopt the declared provider and never resolve authenticating
providers, and a router instance without get_model_list degrades the gate
to a no-op. Tests rebuilt on real Router instances resolving deployment
model_info end to end, plus a full-path test through
async_get_available_deployment
A guardrail modify_response verdict on a streaming request only produced a
proper replacement on /v1/messages: the chat completions and Responses API
translations had no build_block_sse_chunks, so the ModifyResponseException
re-raised and surfaced as an in-stream 500 error frame (or a whole-request
500 in buffered mode) instead of the documented 200 replacement.
Implement build_block_sse_chunks for both OpenAI translations: chat emits a
content delta plus a finish_reason content_filter chunk with real usage;
Responses emits the typed event sequence (standalone via
build_synthetic_response_events pre-stream, or an output-item continuation
under the in-progress response id mid-stream) ending in response.completed.
UsersTable overrides DataTable's default noDataMessage with its own
EmptyState, so the row reads "No users found" rather than "No results".
Assert that, and pair it with the seeded user being absent so the check
cannot pass while the filter silently does nothing.
A non-admin switching an existing key's type between the safe preset
buckets (llm_api_routes, info_routes, and empty = full access) got a 403
from the allowed_routes admin gate, because /key/update, unlike
/key/generate and /key/regenerate, had no carve-out for preset-derived
values. Skip the gate only when both the incoming and the stored
allowed_routes consist entirely of safe presets, so clearing an
admin-set custom route restriction still requires proxy admin.
The batch start told the seed which LiteLLM_SpendLogs rows were its own, but
using it as a hard cutoff also dropped rows another pod had already persisted.
Those rows are only repaid by that pod's own increment, so if it died first the
window row stayed permanently under the recorded spend.
The seed now reads both sums in one scan and takes off this batch's own spend,
flooring at the pre-batch total for the case where its log rows have not landed
yet. Redis payloads keep an empty request_ids so a leader from before the field
was dropped can still merge what it pops during a rolling deploy.
Claude-Session: https://claude.ai/code/session_01QvQzYztinxj8ZuD5YxbVdL
Two assertions were checking the wrong thing. The anchoring tests read
getByRole("listbox"), which resolves to SelectPrimitive.List; that sits at
full content height inside the popup that clips and scrolls it, so the box
overlapped the trigger even when nothing visible did. Measure the popup.
The SSO-ID search expected zero rows, but DataTable renders a "No results"
message row when a filter matches nothing, so the count is one. Assert the
empty state the user actually sees.
GigaChat reports prompt_tokens and total_tokens after subtracting cached
tokens (the docs example is prompt_tokens=1, precached_prompt_tokens=37,
total_tokens=5, so the fields are disjoint, not a subset). Map to the
OpenAI convention by adding precached_prompt_tokens back onto prompt and
total while still surfacing it as prompt_tokens_details.cached_tokens.
xAI states the amount it charged in usage.cost_in_usd_ticks, at 10^10 ticks to
the dollar, and that figure covers tokens and every server-side tool invocation
together. The xAI chat and responses transformations restate it in USD on
usage.cost, the field litellm already carries a provider-stated cost in, and the
xAI cost calculator bills from it the way the perplexity calculator does
Routing it through usage.cost rather than a private field means the streaming
chunk assembler carries it too, and no provider-neutral file has to learn about
an xAI wire field
Only a finite, non-negative amount is trusted, so an endpoint a caller can
point litellm at cannot report a negative amount to subtract from its own
recorded spend, and cannot report a NaN, which Usage stores unvalidated and
which compares false against every budget threshold, disabling enforcement for
the key rather than mispricing one request. Absent a usable figure nothing
changes: the existing token math and the
$5 per 1,000 web search calls fallback both run as before
The web search surcharge is suppressed once the reported total applies, since
that total already covers the search calls
Adds Playwright coverage for the RC checklist items an audit marked
automatable today: Playground to Logs hand-off, public Agent/MCP hub
tabs, team models in the Playground dropdown via a team key, Add Model
with a stored credential, internal user team key creation, a second
admin account, team model deletion, and Presidio guardrail CRUD without
a live sidecar. Seeds e2e-team-keygen with the /key/generate member
permission so the internal user key flow avoids the team-list cache lag
The base added router_metadata to SpendLogsMetadata in #39001 without
updating this fixture, and its CI run never executed logging_testing,
so the job now fails on every branch merged with current staging.
The consolidated popup test only asserted the options never cover the
trigger, so opening above the trigger with room below it, the regression
PR #38554 fixed, would have passed. Split it back into a below-trigger
case and a cramped-viewport case. The header test accepted a single pixel
of vertical intersection; require the refresh control's centre to sit
within the tab row instead.
The migration smoke waited on `getByRole("button", { expanded: false })`
after clicking it. Playwright re-resolves that locator on every retry, so
once the clicked group flipped to expanded it matched the next collapsed
group instead, and the assertion could never pass. Count the remaining
collapsed groups and wait for that count to drop by one.
precached_prompt_tokens is a subset of prompt_tokens (OpenAI cached_tokens
semantics), so map it to prompt_tokens_details.cached_tokens instead of
adding it on top of prompt/total. Emit stream usage from any final chunk
carrying it rather than only finish_reason stop, which dropped tokens for
function_call and length streams. Merge auth metadata into a new dict in
the gigachat router handler instead of mutating the shared parsed-body
cache in place.
The UI e2e suite had a class of assertions that pin how the dashboard is
built rather than what it does, so an ordinary refactor turns them red
without any user-visible change.
Geometry. The auto-router template select had two tests made of pixel
arithmetic plus a data-side="bottom" check, which is Base UI's own
positioner signal. The regression they guard (#38554) is a popup opening
on top of the control that spawned it, so both cases collapse to one
invariant: the options never cover the trigger. It now runs at both
viewport heights and reads the popup as role=listbox. The models header
test compared the tabs and refresh centers within 2px, which a padding
change flips; it now asserts the two share a row.
Structure. The logs drawer test walked xpath=../../.. from a text node
and read collapsed state off chevron icon classes. SectionHeader now
renders a real disclosure button with aria-expanded, and its two copy
buttons carry distinct names instead of both being "Copy". Sidebar group
toggles expose aria-expanded too, so the migration spec can ask for a
collapsed group by state rather than by nesting depth.
Positional lookups. keyRow.locator("button").first(), row.locator("td")
.first() and getByTestId(grid).locator("div").first() all named a
position where they meant an action; they now name the control. Table
scoping moves from "table tbody" to role=row.
Timing. Nine waitForTimeout calls are gone. Every assertion that followed
them already retried to its own timeout, so the sleeps only slowed the
run down.
Both files under tests/users/ were wrapped in test.skip("...", () => {}),
which registers one skipped test and never runs the body, so the four
tests inside had never executed and were written against a UI that has
since changed (the search placeholder is "Search by email…", the ID
filters moved into a drawer, pagination is labelled "Go to previous
page"). Rewritten against the current surface: the suite goes from 104
collected tests to 107.
Left in place deliberately: the chip and dialog-footer data-slot
selectors, because the accessible names they work around live in
components/ui/, which is shadcn CLI-managed and not hand-edited.
* fix(otel): emit cache token counts on OTel v2 LLM spans
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): trim comment in LLMUsage adapter
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): drop casts in LLMUsage cache token adapter
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(deps): bump restrictedpython to 8.3 for GHSA-ffg3-p8fm-mjx2
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
`ChatCompletionFileObject` is in the union `_count_content_list` accepts, but
`file` was missing from its match, so every local count of a Responses
`input_file` raised `Invalid content item type: file`. On
/v1/responses/input_tokens that surfaced as an opaque 500 whenever the model's
provider counting API refused the block and the local tokenizer took over.
Count it the way the module already counts the same thing in Anthropic's
dialect: the filename like a document title, the inline bytes through the
image pricer.
The Responses-to-chat transform dropped the filename OpenAI requires next to
file_data, so a request carrying an inline PDF counted 13 tokens instead of 36
and a real completion through the chat bridge got a 400.
