delete_deployment resolved the outgoing deployment through get_deployment before
popping it, and ran the strategy release before repairing the index maps. Both
halves of that ordering could leave the router inconsistent. A resolution failure
meant the entry left the model_list with its registry slots still held, so the
alias stayed routable and the name could not be reused; a failure inside the
release meant the outer handler returned None with the entry already popped and
model_id_to_deployment_index_map never repaired, breaking every later lookup and
delete until a restart.
upsert_deployment already had this right: it pops, repairs the caches and indices,
and only then releases the slot. delete_deployment now follows the same sequence
and resolves the deployment from the item it just popped rather than through a
lookup that can fail. Releasing the slot is secondary to structural integrity, so
it runs last and a failure there is logged instead of abandoning a removal that has
already happened.
* fix(ui): validate default team values in Default User Settings
The Default User Settings form accepted any free-text team id, and the
proxy persisted it without checking the team exists. New users were then
silently never added to the default team because the consume-time 404
from team_member_add was swallowed at debug level.
Backend: PATCH /update/internal_user_settings now rejects unknown and
duplicate team ids with a 400 naming them, before any persistence or
team budget side effects. Team-add failures in _add_user_to_team now log
at ERROR with user and team ids.
UI: DefaultUserSettings rewritten as a shadcn + react-hook-form + zod
form following the org-settings pattern. The team id free-text input is
replaced with a searchable server-backed team picker, so only existing
teams can be selected; zod blocks empty and duplicate rows. The shared
deriveErrorMessage helper now unwraps the HTTPException detail.error
shape so backend validation errors surface readably in toasts.
* fix(ui): restore read-only view with Edit Settings toggle on default user settings
Parity with the pre-migration form: the tab renders a read-only summary
of the saved defaults, Edit Settings opens the RHF form, Cancel discards
pending edits and returns to the summary, and a successful save returns
to the summary showing the new values. Model sentinel labels in the
summary are derived from ModelSelect's now-exported special values
instead of duplicating the strings.
* refactor(ui): rename MODEL_SELECT_SPECIAL_VALUES_ARRAY to MODEL_SENTINEL_OPTIONS
* fix(ui): move Edit Settings into the card header action slot
The finalize re-run in upsert_deployment keyed off the auto_router/adaptive_router
prefix only, so editing a complexity router with adaptive enabled released its
adaptive_routers entry (and post-call hook) without rebuilding it: complexity
routing kept serving while bandit recording, DB persistence and
/adaptive_router/state went silently dark until the next full reload. Gate the
re-run on a participation predicate that mirrors both arms of the finalize pass,
drop the import that pass no longer uses, and pin the registry helpers with
direct contract tests
Auto-router-family deployments live in two structures: the model_list, and a
pre-routing strategy registry keyed by (model_name, tags). Removing a deployment
dropped it from the model_list without releasing its registry slot, so the re-add
that follows hit the "already exists" guard in _register_pre_routing_strategy and
ignore_invalid_deployments swallowed it. The deployment came out and never went
back, while the DB row and the endpoint response both looked fine. Only a restart
healed it, and under multiple replicas each pod diverged into holding a different
subset of routers.
Removal now releases the (model_name, tags) slot from every strategy registry, in
both upsert_deployment and delete_deployment, guarded on the auto_router/ prefix so
removing a regular deployment cannot evict a router that merely shares its
model_name. Releasing from every registry rather than the first match is what makes
this correct for hybrids: registration is one-to-many, since a complexity router
configured with adaptive is also registered in adaptive_routers under the same key
by the deferred finalize pass. Releasing only the first match left that adaptive
strategy live, so a deleted or replaced alias stayed routable through it.
Adaptive post-call hooks are rebuilt whenever the adaptive registry changes, not
only at the end of set_model_list. The hook set is defined as exactly one hook per
registered adaptive router, so a released router stops recording turns instead of
holding a hook bound to a strategy nothing points at any more.
The swallowed upsert failure is logged at warning instead of debug, which is below
the default log level and left this failure with no observable signal anywhere.
delete_deployment resolves the outgoing deployment before popping it, and a
resolution failure no longer aborts the removal; previously an entry that failed
validation would have been left in the model_list permanently.
delete_model drops its blanket pop across all four registries. That predates this
change and over-evicts: it removes every tag variant registered under the name
while only one is being deleted, and nothing reloads on that path to restore the
survivors. delete_deployment now handles it correctly and tag-scoped, so the
endpoint-level eviction and its helper are removed rather than left to mask it.
* test(e2e): wait for MCP tool discovery instead of racing it
/v1/mcp/server returns as soon as the DB row is written, but the gateway runs
the initialize + tools/list handshake against the upstream lazily, on the first
request that needs it. Every MCP test read tools/list immediately after
registering, so it raced that handshake.
The gateway reports a server it has not discovered yet exactly like a dead one:
it catches the per-server handshake exception and returns an empty tool list.
The tests asserted on a single read, so the race surfaced as "granted key never
saw search_datadog_logs; tools=frozenset()" while a sibling test against the
same upstream in the same run passed.
Add McpClient.await_tool, which polls tools/list to the suite's existing
poll_timeout and returns the qualified tool name, and route the four discovery
sites through it. An unreachable upstream or an unapplied grant still fails, and
the failure now names the last tools/list result.
Refs LIT-4821
* test(e2e): wait for a2a agents to reach the data plane after registration
POST /v1/agents is a control-plane write; the /a2a/{agent_id} routes that serve
the card and run message/send are data plane and only see the agent after the
next DB reload. Every test registered an agent and immediately read its card or
sent it a message, so the first data-plane touch could 404 on the agent it had
just created.
register_agent now waits for the card to become servable before returning, the
same way ProxyClient.create_model waits for a new model, so callers do not each
have to poll. Registration failures skip the wait, leaving the two rejection
tests unchanged. A genuine propagation failure now fails naming the agent id and
the last card read rather than as a bare 404 on whichever /a2a call ran first.
Refs LIT-4821
* test(e2e): wait for presidio guardrails to sync before asserting masking
Registering a guardrail is a control-plane write; the data-plane worker that
serves /chat/completions only picks it up on its next periodic DB sync (~30s), so
the first call after the create ran against a worker with no guardrail and passed
the raw email straight through. The tests asserted on that first call, so they
read in-flight propagation as a PII leak.
Confirmed directly against a live proxy: the same call is unmasked at t=0s and
masked at t=8s, and the presidio analyzer itself correctly returns EMAIL_ADDRESS
with score 1.0 the whole time. The MCP guardrail suite already documents and
waits out this exact sync delay; presidio never got the same treatment.
Poll the call until the placeholder replaces the PII, so the assertions judge the
synced state. A guardrail that never masks still fails, on the last unmasked
content. pre_call and post_call now pass repeatably.
Refs LIT-4821
* test(e2e): drop the presidio logging_only check pending LIT-4841
pre_call and post_call masking both pass once the guardrail-sync wait is in place,
but logging_only left the raw email in the OTEL span's gen_ai.input.messages on
every attempt across a full poll deadline. Keeping an assertion against
known-failing behavior just turns every run red, so the cell is tracked in
LIT-4841 instead.
The registry row stays, so guardrail.presidio.logging_only.masks now reports as an
uncovered gap rather than silently disappearing.
Refs LIT-4821, LIT-4841
* test(e2e): wait for guardrail sync in bedrock, moderation and block-code checks
All three asserted on the first call after registering a guardrail, so they were
served by a data-plane worker that had not synced it yet (~30s DB poll) and read
in-flight propagation as a guardrail that failed to block. Verified directly: the
openai_moderation guardrail lets a flagged prompt through at t=0s and returns
"Violated OpenAI moderation policy" at t=8s.
The reasoning-only responses noted in triage (content=None with reasoning_tokens
set) were a symptom of the same thing, not the cause; these are pre_call
guardrails, so a synced guardrail rejects the request before the model runs.
Add poll_until_blocked to guardrails_client for the two that surface a non-success
status, and poll on the block marker in the block_code_execution check, which
replaces the reply rather than erroring. All eight guardrail tests now pass.
Refs LIT-4821
* test(e2e): drop the openai prompt-cache check pending LIT-4841
Prompt caching never engages through the proxy: cached_tokens is 0 on every
repeat, while the identical payload sent straight to OpenAI reports 3615 cached
tokens on the second call. Pinning prompt_cache_key on the proxy request restores
caching (3328 tokens), so something varying per request is defeating OpenAI's
automatic prefix cache.
That is a product bug with a direct billing cost, tracked in LIT-4841. The
registry row stays, so llm.chat_completions.openai.prompt_cache_5m.nonstream.works
now reports as an uncovered gap instead of failing every run.
Refs LIT-4821, LIT-4841
* test(e2e): drop the responses metadata redis-ttl check
It failed on a Redis read timeout against the stage serverless cache
(berrie-litellm-stage-ieib2i.serverless.use1.cache.amazonaws.com:6379), a
reachability problem this suite has hit before rather than a proxy defect the
assertion can pin down.
The file held only this test. Its other cell,
llm.responses.openai.basic.nonstream.works, is still covered by
test_responses_e2e.py; other.config.responses.metadata_redis_ttl_bounded becomes
an uncovered registry row, taking headline coverage 314/431 -> 312/431.
Refs LIT-4821
* test(e2e): fix passthrough header propagation and openai body, drop the cost check
Three separate problems behind the two passthrough failures.
The header test 404'd because POST /config/pass_through_endpoint is a
control-plane write and the worker serving the route only registers it on its next
config reload; measured at ~18s on a live proxy. Wait for the route to stop 404ing
before calling it. The readiness probe reuses the master key and omits
anthropic-version so polling does not bill a completion per attempt.
The openai passthrough body sent max_tokens, which the gpt-5 family rejects
outright ("Unsupported parameter: 'max_tokens' is not supported with this model").
Confirmed against OpenAI directly: max_tokens 400s, max_completion_tokens 200s.
Passthrough forwards the body untouched by design, so the body was simply wrong.
test_openai_passthrough_nonstreaming_logs_cost still finds no SpendLogs row for
its call_id after the fix, so it is removed rather than left red; the gemini and
anthropic passthrough cost checks still cover that path.
Passthrough suite is 8/8 green.
Refs LIT-4821
Exercises the streaming field-fill heuristics, model_construct fallbacks,
and the get/cancel/delete/list request and response transforms that had no
tests.
litellm_provider_cache_creation_input_tokens_metric only read the
Anthropic-style top-level usage.cache_creation_input_tokens and had no
prompt_tokens_details fallback, unlike its cache-read twin. OpenAI models
that bill prompt cache writes report them only in
prompt_tokens_details.cache_write_tokens, so the counter never fired for
them. Resolve provider cache read/write tokens through a shared helper
that falls back to prompt_tokens_details.cache_write_tokens (canonical)
then cache_creation_tokens when the explicit top-level field is absent,
and give litellm_input_cache_creation_tokens_metric the same fallback for
raw usage dicts that only carry cache_write_tokens
Resolves LIT-4823. An adversarial review reproduced against real Postgres
that a batched increment upsert stalling past the prisma engine timeout
leaves its transaction open on the pooled connection; the retry draws the
same connection, its statements stack into the still-open transaction, and
one commit applies both increment sets while the writer reports success.
httpx.ReadTimeout is exactly that post-send case and every spend writer
retried it.
DB_RETRY_SAFE_ERROR_TYPES (ConnectError only, the failure that proves the
statements never reached the database) is now the single owner of what a
non-idempotent writer may retry. All seven entity and daily spend writer
retry arms and the tool usage flush consume it. DB_CONNECTION_ERROR_TYPES
is unchanged for the idempotent spend-log writer, whose create_many with
skip_duplicates may safely retry the full tuple.
The corruption was reproduced on update_daily_user_spend (seeded 10|100|1,
expected 11|110|2, observed 12|120|3); the new policy tests pin that a
ReadTimeout drops the batch loudly on the first attempt and a ConnectError
still retries.
Anthropic-format requests translate to messages whose content is a list
of part dicts, which the headroom compression service's transforms
silently skip (they only rewrite string content), so compression never
applied to Anthropic client traffic while the guardrail still reported
itself as applied.
Flatten all-text part lists to plain strings for /v1/compress and
restore the original shapes from the response: untouched rows keep
their exact original parts, a rewritten row collapses to one part
carrying the last declared cache_control breakpoint (a breakpoint
caches the prefix ending at its part, so the last one and its TTL
still describe the merged row). Rows with any non-text part are never
flattened, since merging text across a non-text part would move a
later breakpoint to the other side of it; they pass through the
service untouched, matching its own behavior for non-string content.
Flattening and write-back use the shared content_text helpers that
compresr's breakpoint fix also uses.
Resolves LIT-4795
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Same cause as the otel handler test: this file builds its request as a
SimpleNamespace carrying only `state`, and the validation handler now reads
`request.url.path` to pick an error contract, so the fake needs a url
While here, cover what the two existing tests do not. They only exercise the
proxy-wide 422, and the control plane's 400 problem document was reachable only
through the route test, which registers its own copy of the handler in a local
app rather than the real one. Two cases now pin the real handler directly: a
`/management/v1` path returns problem+json with a `detail` string, and paths that
merely resemble the prefix (`/management`, `/v1/management/foo`) keep the 422
shape their callers parse
Both failures are from this branch, not pre-existing
The component allowlist test asserts the gateway and backend route sets union to
the whole app, so any route on neither is a 404 on both pods. Allowlist the
`/management/v1/` prefix on the backend, next to the other control plane
entries, so every resource that moves under it later is covered without a
per-resource edit
The otel handler test builds its request as a SimpleNamespace carrying only
`state`. The validation handler now reads `request.url.path` to decide whether
the caller is on a surface with its own error contract, so the fake needs a url;
a real Request always has one, which is why the handler does not guard for it
The control plane branch returns early, and nothing covered that it still closes
the dangling SERVER span first, so those requests would have leaked a span
apiece. Added a case that pins it; removing the close call fails it
Three fixes from an adversarial review of this branch, each at the owning
seam rather than the report site.
The flush retried DB_CONNECTION_ERROR_TYPES, which includes ReadTimeout.
A ReadTimeout is the committed-but-unacked case: the review reproduced the
engine abandoning the transaction open on the pooled connection, the retry
stacking its statements into it, and one commit applying both increment
sets while the flush reports success. The retry now covers only
ConnectError, the one failure that proves the statements never reached the
database; post-send failures drop the batch with an error log. The
docstring no longer claims an idempotency the pattern does not have. The
same hazard exists in the untouched daily spend writer and is left for its
own change.
get_tool_calls_from_response read choices[0] only, so a tool invoked in a
later choice of an n>1 response earned spend but never reached the rollup,
the index, or the registry. Choice scope is now an explicit parameter:
accounting passes include_all_choices=True because every choice costs
money; guardrails keep the primary-choice default because they rebuild the
primary assistant message. First multi-choice fixtures in the suite pin
both scopes.
maxBarSize=64 had been added to the shared BarChart unconditionally,
resizing every existing consumer. It is now a prop; only the tool spend
charts opt in. The legend flex-wrap changes stay global because clipping
overflow was a defect, not a preference.
The RFC 9457 `type` was `https://docs.litellm.ai/errors/<slug>`, copied from the
standard's own error example. That path is a 404 and there is no docs section
behind it, so every error body shipped a broken link
RFC 9457 only requires `type` to identify the problem type; it encourages, but
does not require, that dereferencing it yield documentation. An https URI makes a
promise we are not keeping, so use `urn:litellm:error:<slug>` instead, which
carries the same machine-readable identity with nothing to resolve. Switching to
an https base later is a contract change for anyone matching on `type`, so that
should wait for pages that actually exist
A test pins the identifier against regressing to an https docs URL, since the
existing assertion built the expected value from the same constant and would have
stayed green whatever it held
`/customer/aliases` shipped two days ago and has not been in a release, so its
wire contract is still free to change. This lands it on the control-plane
contract before that stops being true, since after a release the path, the param
names and the envelope would all need a permanent legacy adapter
The endpoint becomes `GET /management/v1/spend_logs/end_users`. It is a facet,
the distinct values one column takes over a filtered query on a resource, not an
entity collection; naming it after `customers` implied it listed the end-user
table when it actually reads spend logs, which is a different row set. Serving it
under the parent resource means its filters are the parent's filters, so the
dropdown offers exactly the values the logs table can show without two endpoints
having to keep agreeing on that
Contract changes: `size` becomes `page_size`, `search` becomes `q`, the window
moves from flat `start_date` / `end_date` to `filter[startTime][gte]` / `[lte]`,
and the body becomes `{data, meta, links}`. Unknown query params are now a 400
rather than being silently dropped, because an ignored filter over-returns data.
Errors are RFC 9457 problem documents on this prefix only; every other route
keeps the shape its callers already parse
`links` is what makes the rest deferrable. The dashboard hook follows the
server's `links.next` instead of computing `page + 1`, so moving this to cursor
pagination later changes the links and nothing the client does. That matters
because the inner scan is a sliding window, so offset paging can currently skip
or repeat an end user across pages; the fix is a follow-up, and the hypermedia
means it will not be a breaking one
Cursor mode, `sort`, `include`, ETag / `If-None-Match` and the generic `ListSpec`
framework are all deliberately out of scope here. They are additive or internal,
so none of them needs to beat the release
GET /v1/tool/spend served the Cost Optimization card with two raw queries
over LiteLLM_SpendLogToolIndex x LiteLLM_SpendLogs on every dashboard load;
the totals query's driving scan was all of SpendLogs in the window. Both
per-request tables reach 1M+ rows at customer scale, so the card cost
O(traffic) per view and had to be capped at 30 days.
The index writer also mined proxy_server_request.tools, i.e. tools DECLARED
in the request body, attributing each request's full spend to tools that
never ran; and all non-MCP mining ran against payload fields that are '{}'
unless store_prompts_in_spend_logs is enabled, so non-MCP coverage silently
depended on a privacy setting.
Now the spend writer builds a ToolUsageTransaction at request time from
invoked tools only, resolved by the shared get_tool_calls_from_response
normalizer so every response surface (chat completions, Responses API,
Anthropic Messages) is covered; the tool registry's response arm delegates
to the same owner. Transactions queue beside the spend-log queue and the
flush job writes index rows plus a new LiteLLM_DailyToolSpend rollup
(date, tool_name PK) in one transaction, retrying connection errors with
backoff (a failed batch commits nothing, so the retry cannot double-count)
and dropping the batch with an error log on anything else.
The endpoint aggregates in SQL: by_tool is the top TOOL_SPEND_TOP_TOOLS
tools by spend via group_by and daily covers only those tools, so the
response is bounded by days x TOOL_SPEND_TOP_TOOLS regardless of range or
tool-name cardinality; the 30-day clamp is gone. total_spend is dropped
from the response; it was never rendered and its deduplicated semantics
are not computable from a rollup. Spend-log retention deliberately does
not touch the rollup, so tool spend history outlives per-request rows.
* fix(guardrails): resolve judge_model credentials via Router in llm_as_a_judge
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): wire llm_router into DB-backed judge guardrail init paths
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(guardrails): assert patch endpoint forwards llm_router to sync
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(guardrails): resolve judge Router lazily and fix wildcard/alias dispatch
Resolve the proxy Router at judge-call time via an injected provider instead of
capturing it at construction, so a DB-backed judge guardrail created before the
Router exists no longer captures None permanently. Select the Router path with
router.get_model_list(model_name=judge_model) so wildcard routes and
model_group_alias keys resolve, not just literal deployment names. Isolate the
judge call from user-traffic routing with num_retries=0 and fallbacks=[].
Revert the llm_router threading through the DB sync/reinit/create/approve/patch
paths since the lazy provider makes it unnecessary. Replace mocked-Router tests
with real Router coverage for plain deployments, model_group_alias, and wildcard
routes, plus lazy per-call resolution.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): harden judge verdict parsing and guard proxy import
Strip markdown fences and surrounding prose before json.loads so fencing-prone
judge models evaluate instead of failing open, guard the proxy_server import in
_default_router_provider so an unimportable proxy falls back to the SDK, and
snapshot/restore global callback lists in the DB-path judge registry tests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): reject non-object judge verdicts instead of failing open as success
* fix(guardrails): route hidden model_group_alias judge models through the Router
---------
Co-authored-by: milan <milan@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: yucheng-berri <yucheng@berri.ai>
Vertex passthrough classified any target URL containing "stream" as a streaming
request. `:streamRawPredict` carries that substring, so a unary Claude-on-Vertex
call whose body omits `stream` was routed through the streaming logging path.
That path never consults the response content-type, so a complete
`"type": "message"` JSON body was handed to the Anthropic SSE chunk parser,
which recognises none of it; the spend log recorded 0 prompt tokens,
0 completion tokens and zero cost
Streaming for the rawPredict family now comes from the request body, which is
what the Anthropic Messages contract uses for those endpoints. The
generateContent family keeps its URL signal because the Gemini REST body has no
`stream` field, and `?alt=sse` is still appended for every request that is
classified as streaming, so Gemini framing and its usage parsing are unchanged
Both passthrough streaming predicates read `.get("stream")` off a body that is
only annotated as a dict; `_read_request_body` returns whatever the JSON parser
produced, so an array body raised AttributeError. The two predicates are now one
owner that answers False for any non-object body, which covers the vertex,
mistral, anthropic, vllm and azure passthrough routes
Cache web identity STS credentials in the shared IAM cache (restores the
pre-v1.85.0 behavior removed by #27125) and cap the Google OIDC token cache
TTL at the token's own exp claim minus a 60s margin, never caching an
already-expired token
* fix(e2e): stop the cache-settings test from persisting a degraded Redis config
TestCacheSettings.test_update_persists_cache_backend_to_get read the live cache
settings and wrote them back, intending a no-op. Its capture modelled only
type/host/port, so on a TLS cluster the write-back silently dropped `ssl` and
`redis_startup_nodes`.
That is not recoverable on its own. `/cache/settings` persists what it receives
into LiteLLM_CacheConfig, that row outranks the YAML `cache_params`, and
init_cache_settings_in_db re-applies it on a timer, so a restart does not clear
it. The proxy ends up driving a TLS-only cluster endpoint as a plaintext
standalone node and every Redis call blocks to socket timeout.
On the affected deployment that took out rate limiting entirely (the v3 limiter
is a Lua script on Redis with no DB fallback), Redis-only budget levels (tag,
per-model, team-member, per-window), spend tracking, `ResetBudgetJob` (which
self-starved at 54 skipped runs per 15 min), and `ProxyConfig.add_deployment`,
whose last statement syncs guardrails and never ran. 60 of 72 failures in one
run traced back here.
The settings blob is now round-tripped verbatim via a RootModel over an
exhaustive value union, so a subset cannot be written. Two guards make a
regression fail loudly at this test instead of silently downstream:
- refuse to write when GET reports redis_type=cluster but omits
redis_startup_nodes, which is the exact precondition for persisting a
downgrade. GET resolves the stored row overlaid with REDIS_* env and never
reads YAML, so a cluster configured only in YAML cannot round-trip here
- compare /cache/ping before and after, so a write that breaks connectivity
fails this test rather than every suite that follows
The underlying product defect is filed as LIT-4816: GET cannot express the
effective config, and a partial POST is allowed to downgrade transport. This
change only stops the suite from triggering it; the Admin UI can still do so.
basedpyright clean (0 errors) under the e2e gate.
* fix(e2e): scope the bedrock guardrail per request and send OpenAI's current token param
Two failures that had nothing to do with the guardrail or route under test.
create_bedrock_guardrail registered with default_on=True, which applies the
guardrail to every request the proxy serves. The upstream ApplyGuardrail call was
answering 403, and that came back to unrelated traffic as
`403 Bedrock guardrail request failed`, failing three a2a tests and a passthrough
headers test alongside the bedrock one. The harness already supports the
per-request `guardrails` selector, so the guardrail is now registered opted out of
default_on and selected by the test that wants it. A broken upstream guardrail
fails its own test instead of whatever else is running.
Note this only contains the blast radius; the 403 itself still needs the
bedrock:ApplyGuardrail permission (or a valid guardrail identifier) on the
deployment, so test_bedrock_pre_call_blocks_harmful_prompt can still fail on its
own until that is sorted.
The OpenAI passthrough body sent `max_tokens`, which newer models reject with
"Unsupported parameter: 'max_tokens' is not supported with this model. Use
'max_completion_tokens' instead." Passthrough forwards the body untranslated, so
drop_params does not apply and the body has to satisfy OpenAI's contract
directly. vllm_chat keeps max_tokens, which vLLM accepts.
basedpyright clean (0 errors) under the e2e gate.
* fix(e2e): drop the pinned a2a api_key that broke every message/send
#34512 pinned `api_key="os.environ/ANTHROPIC_API_KEY"` on the a2a bridge agent.
The a2a bridge forwards the agent's litellm_params straight into
litellm.acompletion() without expanding "os.environ/" indirection, so that literal
string was sent upstream as x-api-key and every message/send failed with
`AnthropicException - {"type":"authentication_error","message":"invalid x-api-key"}`.
Omitting api_key restores the normal provider resolution: litellm reads
ANTHROPIC_API_KEY from the proxy's own environment for this provider, which is what
the agent-owner flow depends on and what the suite did before #34512.
Verified against a live proxy, same agent shape each time:
api_key omitted -> message/send 200
api_key "os.environ/ANTHROPIC_API_KEY" -> message/send 500 invalid x-api-key
api_key <literal key> -> message/send 200
and the key itself is valid (direct call to api.anthropic.com returns 200), so this
was indirection that never got expanded rather than a bad credential.
This accounts for four failures (test_semver_protocol_version_registers_and_serves,
test_message_send_runs_completion_bridge, test_pinned_v0_3_serves_flat_message_shape,
test_pinned_v1_0_serves_nested_message_shape). They were previously reported as
`403 Bedrock guardrail request failed`, because a default_on Bedrock guardrail
short-circuited the request before it ever reached the bridge and hid this.
The bridge silently ignoring "os.environ/" in agent params is a product defect in
its own right, filed separately; anyone configuring an agent credential that way
through the UI hits the same wall.
basedpyright clean (0 errors) under the e2e gate.
* test(e2e): make the load suite less aggressive against a shared proxy
750 users at spawn rate 50 saturated the request path hard enough to distort the
latency-sensitive suites sharing the same proxy, and it spends real provider money
at that rate. Drop to 200 users at spawn rate 20.
The RPS floor moves with the user count rather than staying put, so the assertion
keeps its meaning instead of becoming a formality: 355 RPS over 750 users is
~0.47 RPS/user, and 90 over 200 holds that same per-user expectation with a
similar pass margin. A request-path regression still trips it.
All four knobs stay env-overridable (E2E_LOAD_USERS, E2E_LOAD_SPAWN_RATE,
E2E_LOAD_DURATION_SECONDS, E2E_LOAD_MIN_RPS) for a deliberate load run.
Note the recorded failure for this test was "no requests completed in 60s", which
was the gateway wedged on unreachable Redis rather than a throughput regression;
this change is about not perturbing its neighbours, not about that failure.
* fix(e2e): make the reasoning-tokens assertion exercise a request that reasons
test_openai_chat_reasoning_reports_reasoning_tokens asked "A train travels 60 miles
in 1.5 hours. What is its average speed in mph?" at reasoning_effort="low", then
asserted reasoning_tokens > 0. The model answers that directly without reasoning, so
0 is correct behavior and the assertion was testing the model's discretion rather
than litellm's reporting.
Verified against a live proxy on a dedicated openai/gpt-5.6 deployment, matching how
the test provisions its model:
reasoning_effort=low, one-step arithmetic -> reasoning_tokens=0
reasoning_effort=high, the prompt used here -> reasoning_tokens=114
Raised to high effort with a prompt that requires a proof plus a search, so the
field under test is actually populated and the assertion fails only if litellm stops
surfacing it.
While confirming this I also checked prompt caching, which needed no change:
cached_tokens comes back 3615 of 3618 prompt tokens on a repeated large prefix
against a dedicated deployment. An earlier reading of 0 was an artifact of probing a
fan-out alias whose requests land on different deployments, not a caching defect.
* test(e2e): skip the files-list test while LIT-4820 is open
GET /v1/files does not include a just-uploaded file. The upload returns 200 and
GET /v1/files/{id} resolves it, but the listing never contains it: the returned set
stays fixed at 27 entries whose newest created_at is roughly ten hours older than
the upload, on both the managed (/v1/files?model=) and provider-scoped
(/openai/v1/files) routes. Polled for 40s, so not an eventual-consistency window.
Filed as LIT-4820. Skipping keeps a known, ticketed product bug from holding the
suite red and masking a new regression somewhere else in the same test.
The assertion is left exactly as it was on purpose. It encodes the contract we
actually want, that a file retrievable by id is also enumerable, and anything that
lists files (a UI picker, cleanup tooling that lists then deletes and would
therefore leak provider-side files) depends on it. Relaxing it to get green would
delete the signal. The skip reason says so and links the ticket, and the ticket
records that removing this marker is part of its definition of done.
Matches the existing pattern in this file, where test_unified_file_and_batch_create
skips with a reason citing LIT-3266.
While skipped, the registry cell llm.files.openai.list.nonstream.works has no
passing covering test, so files-list coverage reports as uncovered rather than
passing, which is the honest state.
* fix(e2e): parse Sentinel node lists in the cache-settings model
The value union covered scalar lists and lists of mappings, but not lists of
lists. `redis_startup_nodes` holds host/port mappings while `sentinel_nodes` holds
positional pairs (CACHE_SETTINGS_FIELDS documents `[['localhost', 26379]]`), so on
a Sentinel deployment pydantic rejected the response:
sentinel_nodes.list[dict[str,...]].1
Input should be a valid dictionary [input_value=['localhost', 26380]]
The round-trip test reads GET /cache/settings before it writes anything, so that
rejection failed the test at the read, before any assertion ran. A Sentinel
deployment would have looked like a broken cache-settings route rather than a
model too narrow to parse a documented shape.
A list element may now be a scalar, a list or a mapping, which covers both node
shapes without special-casing either and tolerates a heterogeneous list instead of
rejecting the whole response.
Adds TestCacheSettingsModel, harness-level with no `e2e` marker so it runs without
a proxy, covering all four backend shapes (cluster mappings, sentinel pairs, plain
node, url mode with a null discrete field) plus transport() key selection.
Confirmed it fails on the previous union and passes on this one:
old union -> 1 failed, 4 passed (the sentinel case)
new union -> 5 passed
* test(e2e): remove the cache-settings round-trip test
The test could not fail for the thing it claimed to test, and could break the
deployment it ran against. Both halves of that are worth stating.
It read the live settings, wrote back identical values, and asserted the read-back
matched. If POST /cache/settings were a complete no-op that returned 200 and touched
nothing, GET would still return the values read a moment earlier and the test would
pass. It verified that GET is stable, not that the route persists anything.
Against that, /cache/settings persists what it receives into LiteLLM_CacheConfig,
that row outranks YAML cache_params, and init_cache_settings_in_db re-applies it on a
timer. A write that omits ssl or redis_startup_nodes converts a TLS cluster into a
plaintext standalone client and every later Redis call blocks to socket timeout. On
2026-07-25 that failed 60 of 72 tests in one run: rate limiting stopped enforcing,
Redis-only budgets admitted billable over-budget spend, ResetBudgetJob self-starved,
and guardrail sync never ran.
Guarding the previous shape was not sufficient. Writing the blob verbatim plus a
cluster precondition and a /cache/ping check narrowed the hazard but did not remove
it, because GET cannot express the effective config: it resolves the stored row
overlaid with REDIS_* env and never reads YAML. On a fresh deploy it cannot see
YAML's ssl to echo back, so a TLS non-cluster deployment could still have a row
written that drops it. No round-trip through this route is safe on a shared proxy.
Removed with the models and helpers it owned, and TestCacheSettingsModel with them
since it existed only to protect that parsing.
The registry row mgmt.cache_settings.update.happy_path stays, now carrying the
rationale for why it is deliberately uncovered and what a safe test would require
(an isolated proxy, or LIT-4816 fixed so a partial write cannot downgrade
transport). Coverage therefore reports this cell as a gap, which is the honest
state. Collector passes --strict; the module still collects 11 tests.
Anthropic cache_control breakpoints are positional: each one caches the
prefix ending at the part that carries it. Compresr flattened every text
part of a message into one string and wrote the compressed result back
into the first text part only, which dropped every later breakpoint and,
when a non-text part sat between text parts, moved the trailing text to
the other side of it.
The positional invariant now has one owner. guardrail_hooks/content_text.py
holds content_to_text alongside is_all_text_parts and
merge_rewritten_text_parts, so a compressed string is only ever written
back over a contiguous run of text parts, and the merged part carries the
last declared breakpoint and its TTL.
Compresr consumes that owner at both ends: _select_targets no longer
selects a row holding a non-text part, and _replace_text_in_content
returns such a row unchanged rather than merging across it. Rows whose
content is a plain string are unaffected.
Mixed rows therefore stop being compressed, which is a deliberate trade;
no single-string write-back can preserve a breakpoint across a non-text
part, so the alternative is silently caching a different prefix than the
caller configured.
* fix(auth): route JWT default-team into memberships instead of the create payload
JWT auto-provisioning (get_user_object with user_id_upsert) merged
litellm.default_internal_user_params verbatim into the Prisma user create,
including a teams key. When a default team is configured through the Admin
UI it is stored as a list of NewUserRequestTeam objects, but the user
table's teams column is String[], so the create raised a Prisma type error
and every JWT-authenticated request 401'd with the user never created.
Mirror the /user/new path: strip teams (and available_teams) out of the
create payload, then route the configured default team through
check_if_default_team_set / add_new_user_to_default_team so provisioned
users get real membership rows. Reuse the synthetic PROXY_ADMIN
UserAPIKeyAuth pattern already used by the team-upsert path to satisfy the
membership permission gate, and import the helpers lazily to avoid the
auth_checks <-> internal_user_endpoints import cycle.
* fix(auth): propagate max_budget_in_team when adding users to default teams
* fix: use pipe union instead of Optional for UP045 budget
* fix(proxy): enforce global max_budget against the resettable proxy budget row
The global proxy budget check compared litellm.max_budget against
SUM(spend) from the MonthlyGlobalSpend view, whose window is hardcoded
to a trailing 30 days. litellm.budget_duration was stored and reset on
a user row that enforcement never read, and startup budgeted the admin
user's own row (default_user_id) instead of the litellm-proxy-budget
aggregate row the spend writer increments per request. Net effect: 1d,
7d and 30d all behaved as a trailing 30 day cap that never reset on the
configured duration.
Startup now upserts the budget onto the litellm-proxy-budget row (and
zeroes lifetime accrual when first putting a row on a reset schedule),
enforcement loads global spend from that row, and ResetBudgetJob drops
the cached global spend accumulator when it resets that row so the cap
unblocks immediately after each window.
Fixes https://github.com/BerriAI/litellm/issues/31292
* refactor(proxy): address review nits on global proxy budget fix
Drop the redundant litellm_proxy_budget_name parameter from
_upsert_proxy_budget_with_reset_at_backfill; its only caller always passed
LITELLM_PROXY_BUDGET_NAME, and any other value would write the budget to a
row enforcement never reads.
Introduce GLOBAL_PROXY_SPEND_CACHE_KEY in constants.py and use it at every
site that previously built the key from litellm_proxy_admin_name (auth
loads, spend-writer increments, startup warm, reset-job invalidation), so
the reader and invalidator can no longer drift apart. The literal key value
is unchanged. Also drop the now-pointless litellm_proxy_admin_name
parameter from _warm_global_spend_cache and the proxy_server import from
the reset-job helper.
