The tool_search x bedrock_invoke cell only ever probed the first turn, so
nothing in the suite has sent a server_tool_use block back to a provider.
Every turn of a real Claude Code session after the first carries the
server_tool_use and tool_search_tool_result blocks the previous turn
produced, and that path was uncovered.
Adds probe_tool_search_multiturn, which takes the real assistant turn
back, answers any client-side tool_use with the id the model actually
emitted, and replays the whole thing as history with the tools still
declared. The assertion refuses to go green unless both server-tool
blocks made it into the replayed history, so a first turn truncated at
max_tokens reads as a failure instead of a vacuous pass.
The replay assertion's red paths never run in a green cell, so they get
markerless harness tests of their own alongside the existing
_builder_unit_tests tree.
No production code.
* fix(proxy): cache tag-name registry so unregistered request tags skip Postgres
Request tags are free-form attribution labels, so most have no LiteLLM_TagTable
row. get_tag_objects_batch never cached that absence: every tagged request ran
a find_many that came back empty, and under Prisma pool contention those
per-request queries queued for minutes inside user_api_key_auth.
Cache the bounded set of registered tag names under one aggregate key with the
management-object TTL. Uncached request tags are filtered against it before any
per-tag DB fetch, so unregistered tags cost zero DB reads on a warm path. An
empty registry is cached as a valid answer; DB errors are not cached and fall
back to the per-tag lookup; tables past TAG_REGISTRY_MAX_SIZE cache an overflow
sentinel that disables filtering. Tag create/update/delete endpoints now evict
the registry and per-tag keys and publish cross-worker invalidation (they
previously evicted nothing). The per-tag write-back also gains the management
TTL it was missing, and the hand-built tag:{name} key strings are replaced with
a shared builder.
* fix(proxy): skip per-request end-user DB reads via restricted-id registry
Every request carrying a user id ran get_end_user_object, and with high-cardinality
auto-created end-user rows (hundreds of thousands of ids, all restriction fields
NULL) the per-pod cache missed on nearly every request, so each one paid a Postgres
find_unique that queued behind the Prisma pool during background-job bursts. True
misses were never cached, and unknown ids paid the read twice per request.
Cache the bounded set of end-user ids that carry any restriction (blocked, budget,
region, default model, or object permission) under one aggregate key with the
management-object TTL. When an id misses the per-id cache and is absent from a
usable registry, get_end_user_object returns None with zero DB reads; restricted
ids keep today's fetch-and-cache path. The skip is bypassed whenever
litellm.max_end_user_budget_id is set (default budgets make unrestricted rows
behaviorally distinct from missing rows), validate_end_user_id_in_db is on
(existence checks need the row), or the token carries end_user_max_budget from
custom auth (the row's recorded spend seeds the budget counter). Empty registries
cache as a valid answer, DB errors are never cached, and oversized tables cache an
overflow sentinel that disables filtering. Customer create/update/block/delete now
evict the registry and per-id keys and publish cross-worker invalidation (they
previously evicted nothing), and the per-id write-back gains the management TTL it
was missing so Redis entries no longer live forever.
* refactor(proxy): single generic registry loader with error sentinel and single-flight
Code review follow-ups on the two registry caches. Registry DB errors now cache
the overflow sentinel for a short REGISTRY_ERROR_NEGATIVE_CACHE_TTL window and
log at warning, so a degraded Postgres stops paying the failing registry scan on
every request on top of the per-id fallback. Cold registry loads are single-flight
per worker behind per-registry locks with a recheck after acquire, so a TTL expiry
no longer fans out one full-table scan per in-flight request. The tag and end-user
loaders collapse into one _load_bounded_registry with per-entity fetch closures,
and the triplicated evict-then-broadcast protocol becomes one evict_and_broadcast
helper beside publish_auth_cache_invalidation, shared by the tag, customer, and
project eviction paths.
* chore(lint): suppress fail-safe registry excepts and ratchet BLE001 budget
* docs(proxy): trim registry cache commentary to single-line why docstrings
* fix(lint): move tag fetch return to else block to satisfy TRY300 budget
Bedrock invoke /v1/messages streaming reports cache_read_input_tokens and
cache_creation_input_tokens on message_stop.usage while attaching
amazon-bedrock-invocationMetrics to the same chunk. The stream decoder
rebuilt that chunk's usage block from inputTokenCount/outputTokenCount
alone, which exclude cache reads and writes, so the cache breakdown was
destroyed before _promote_message_stop_usage could surface it and cache
tokens were billed at $0. Merge instead of replace, and also map
cacheReadInputTokenCount/cacheWriteInputTokenCount when Bedrock reports
the cache itemization inside the invocation metrics.
Co-authored-by: Brian Cox <3924351+brian5021@users.noreply.github.com>
Azure rejects the legacy `max_tokens` key for the whole gpt-5 name family, but
`AzureOpenAIGPT5Config.is_model_gpt_5_model` deliberately excludes `gpt-5-chat*`
so those deployments fall through to `AzureOpenAIConfig`, which sends `max_tokens`
verbatim and gets a 400 back on every request that carries it, `/health` probes
included.
One predicate was answering two independent questions. Split it: the new
`AzureOpenAIConfig.requires_max_completion_tokens` covers the whole gpt-5 name
family and drives only the rename, while `is_model_gpt_5_model` keeps keying
reasoning_effort, the temperature clamp and the dropped penalties off the
reasoning question, so #13781 stays fixed.
Adds an opt-in operator allow-list, litellm_settings::bedrock_request_metadata_fields, that forwards LiteLLM key, team and end-user identity plus client spend_logs_metadata into Bedrock request metadata so Bedrock spend can be grouped in AWS Cost Explorer.
Covers all three Bedrock surfaces: the Converse body requestMetadata field, and a signed X-Amzn-Bedrock-Request-Metadata header on Invoke chat completions and on Invoke /v1/messages, where the header is the only viable leg.
The resolver reads both metadata variable names, reserves the whole user_api_key_ prefix against caller-supplied keys, caps the client slot budget explicitly at 16 minus the reserved count, and drops rather than rejects auto-injected values that violate Bedrock constraints. Caller-supplied requestMetadata keeps its existing 400 semantics.
The request-metadata field and header are proxy-owned whenever forwarding is enabled. A caller-supplied value, reachable through the generic extra_headers passthrough, is dropped unconditionally and compared case-insensitively, and is replaced only by the proxy's own value, so identity in the AWS billing record cannot be forged. Absence of a resolved value still means absence on the wire rather than a fallback to the caller's. The guardrail headers keep their existing no-displace behaviour.
* fix(guardrails): scan text on /guardrails/apply_guardrail for Azure Content Safety
The two Azure Content Safety guardrails never implemented apply_guardrail, so the
endpoint fell through to the base no-op and answered 200 with the caller's text
echoed back, having scanned nothing.
Implementing that method also flips the proxy's unified-vs-native dispatch, which
would move request traffic off these guardrails' own hooks. Add an opt-out that
keeps every lifecycle event on the native hooks, so only the endpoint changes.
* test(guardrails): cover the remaining native-hook opt-out dispatch sites
Adds regression tests for the parallel post-call path, the MCP post-call hook, and
the policy engine step, so every read of the opt-out flag fails when removed.
Callers can opt into the provider's raw operation response on /v1/ocr with the x-req-format: native header (or req_format in the body) while page-based cost tracking keeps reading usage_info off the normalized response.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(ui): link key info header to its user, creator, team, and organization
The key info page showed the owning user and creator as plain text and never surfaced the team or organization at all, so walking from a key to its parent entities meant copying ids into other pages. The header now renders User and Created By as links to the user detail page, and gains a far-right column with Team and Organization links (alias when known, id otherwise). Client-side navigation logic shared by BadgeLink and IdentityCell moves into a reusable EntityLink so all entity links behave the same
* test(ui): mock next/navigation in VirtualKeysTable test
KeyInfoView now renders EntityLink, which calls useRouter, so the table test that opens the key detail needs the app router mocked
* Add default model pin to complexity router UI
A complexity router's default model was only ever derived from the tiers, so
operators had no way to point the fallback at a model that is not first in the
Simple or Medium tier. Adds a Default Model select that records an explicit pin.
The pin is stored in complexity_router_config.default_model, which the backend
already reads, and mirrored onto complexity_router_default_model on save. Both
paths resolve through one helper that mirrors init_complexity_router_deployment:
a pin wins, otherwise MEDIUM or SIMPLE. Recording the pin in the config keeps it
distinguishable from a derived value, so a pin that happens to match the tiers
survives a round trip instead of being read back as tier tracking.
The edit modal only requires one non-empty tier, so a router with models in
COMPLEX alone can reach save with nothing the backend would pick. That now
blocks with an inline message rather than saving a router that raises at init.
* fix(UI): probe the pinned default model in the auto router connection test
The connection test built its targets from the tiers and the embedding model
only, so a Default Model pin outside every tier was never reached and a green
result could hide an unreachable default. model_info_view had already hand
rolled the dedupe and append locally, so the rule moved into
buildAutoRouterTestTargets and both call sites now share it.
* fix(ui): mirror backend precedence when resolving a complexity router default
The edit modal only recognized a pin stored in complexity_router_config.default_model,
so a router whose default lived solely in litellm_params.complexity_router_default_model
lost it on the next save. That field cannot be trusted outright either: before this PR
every save wrote a tier-derived value into it, so treating any value as a pin would
freeze legacy routers away from their tiers. Hydration now takes the config marker as
authoritative and falls back to litellm_params only when it diverges from what the tiers
alone derive, which is only reachable through an external API or config write.
Test Connection had the mirror-image bug: it fell back to complexity_router_config.default_model,
a UI-only marker init_complexity_router_deployment never reads, so it could probe a model
the router would never call. It now follows router.py exactly: litellm_params, else pure
tier-derivation.
Also reword a tooltip that hardcoded the Default Model select's position on the page, and
document the dual write and the create-vs-edit validation asymmetry.
Shape detection and block normalization sat in the generic batch layer, which
let batch and live parsing of the same wire format drift apart. Both now live on
AmazonConverseConfig as is_converse_usage_shape and usage_from_batch_output, so
batch_utils asks the provider adapter rather than knowing Bedrock's field names.
Adds direct coverage for the shape predicate, the completion of an incomplete
block, cache-count inflation, and the streaming usage event that shares the
public transform. Drops the narrative banner from the batch tests.
- decode upstream first frame as utf-8 instead of ascii
- reject model-restricted keys at connect to match HTTP model enforcement
- log the actual request path for /openai_passthrough traffic
Embeddings rows were identified by body shape (has `input`, no
`messages`/`prompt`), which also matches a `/v1/responses` batch row
and reserved zero output tokens for it -- letting a project caller run
large Responses generations against a quota-limited model without
consuming OTPM. Classify embeddings by the row's own `url` instead,
and read `max_output_tokens` as a Responses output cap alongside
`max_tokens`/`max_completion_tokens`.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resolves conflicts from the upstream merge and addresses the Veria-AI
review comment on this PR: batch rows could bypass a project's
per-model ITPM/OTPM quota when the batch's file-bound/routing model
had no quota configured. Charges each row's own model against its own
project quota instead of only the routing model's, and fixes rate
limit error messages to attribute the correct model via a new
descriptor_value field on RateLimitStatus/AtomicCounterMeta. Also
re-syncs the ruff-strict, type-discipline, and basedpyright budgets
against the correct (non-stale) merge base.
Co-authored-by: Cursor <cursoragent@cursor.com>
Every bedrock batch output line went through the Anthropic usage parser, which
reads snake_case input_tokens/output_tokens. Converse-family models (Nova and
friends) report camelCase inputTokens/outputTokens, so their usage came back
0/0/0 and the batch billed $0 despite real token consumption.
Usage is now selected by the shape of the payload: a Converse-shaped block goes
through the same transform the live Converse path uses, so a batch and an
equivalent non-batch call agree on tokens, including cache reads and writes.
Anthropic-shaped bedrock output is unchanged.
A shape neither parser understands (an InvokeModel-native payload from Titan,
Cohere, or Llama, which name their counts differently again) still reads zero,
but now warns with the keys it saw instead of silently billing $0.
Exposes the Converse usage transform as public, since batch parsing is a second
legitimate caller; that also removes the private-member access invoke_handler
was already making.
Replace Any-typed seams with real types in files carrying the highest
remaining reportAny/reportExplicitAny density after #34745: the proxy
server and its utils, the router, the streaming handler and chunk builder,
litellm_logging, the redis cache, the MCP db/tool-registry/spend-writer
layer, the anthropic pass-through adapters and guardrail translation, the
lasso and presidio guardrail hooks, the azure_ai agents handler, the
management endpoints (keys, users, ui_sso, model access groups, config
override, MCP, projects), the responses MCP handlers, response polling
background streaming, and the containers and vector stores mains
No casts, no type: ignore, no noqa, no new suppressions, and no Any
annotations that were not already at base. Whole-tree basedpyright:
reportAny 14,610 -> 14,009, reportExplicitAny 5,100 -> 4,780, total
144,743 -> 143,471, with no rule increasing repo-wide or in any file.
Budgets ratcheted: basedpyright -1,272 across 48 rules, ruff-strict -85,
type-discipline -37
