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.
Live verification against a real OpenAI batch showed per-request
failures (e.g. a rejected param) land in error_file_id, never in the
output file, so failed_requests silently undercounted them (0 instead
of the real 1). _handle_completed_batch now also fetches error_file_id
when present and folds its line count into failed_requests.
* 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>
Batch retrieval already computed cost/usage on completion, but silently
dropped reasoning tokens and never counted per-line success/failure.
Adds BatchCostUsageResult (replacing bare cost/usage/models tuples) with
successful_requests/failed_requests, and threads reasoning_tokens through
the aggregated Usage. Both surface on SpendLogs the same way batch_models
already does.
Replace _attach_budget_limits_usage, which rewrote the caller's key_info dict, with _budget_limits_with_usage returning a new list. Callers assign the result once. Keeps the response shape and spend-counter read path identical while following the repo's no-mutation rule.
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.
Address the Greptile review on #36397. The earlier commit only ratcheted
the LIT010 budget; it never applied the annotations, so
duplicates_in_one_item and duplicates_across_items were still bound
without a Final declaration (LIT010) and the first fixture line was at the
120-char ceiling. Annotate both with `: Final` and wrap the long literal.
RED -> GREEN: check_type_discipline flagged both vars LIT010 before ->
LIT010 gone after (file total 551 -> 549, LIT002 unchanged at 953);
test_calculate_web_search_requests_counts_unique_queries still passes.
Gemini 3 per_query grounding is billed per unique search query the model
executes, ignoring empty queries. _calculate_web_search_requests summed every
non-empty webSearchQueries string across grounding metadata items, so repeated
queries within a request inflated web_search_requests and overstated cost. Count
distinct non-empty queries across items instead.
Fixes#36377
- 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
ResponseAPIUsage.parse_cost already flattens Perplexity's
usage.cost.total_cost dict down to a float before it reaches the
perplexity cost calculator, so the isinstance(cost_info, dict) check
was always False on that path. Every Responses-mode Perplexity model
was silently falling back to manual token-rate calculation and
recording $0 spend whenever static per-token rates were missing.
`/health` already stripped `api_key` from each deployment row via
`ILLEGAL_DISPLAY_PARAMS`, but `extra_headers`, `headers`, and `aws_session_token`
were never added to that list, so `GET /health` leaked provider credentials
(Azure `api-key`, Google `x-goog-api-key`, Bearer tokens, AWS session tokens) in
plaintext to any caller, even without a master key.
Add those three fields to `ILLEGAL_DISPLAY_PARAMS` so `_clean_endpoint_data()`
omits them for all callers, matching how `api_key` is already handled.
Fixes#36898
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>
