* fix(azure_sentinel): split batches under the 1MB ingestion cap and keep undelivered records queued
Azure Monitor rejects any Logs Ingestion body over 1MB with a 413. The Sentinel logger
posted the whole queue as one body and cleared it in a finally block, so an oversize
batch, a transient 5xx, or a failed token call dropped every queued record, and records
logged while a send was in flight were cleared with it. Both the standard and the audit
queue share the sender.
Move Datadog's proactive size split and 413 halving into a shared helper,
litellm/integrations/batch_utils.send_batch_with_413_split, and route Sentinel through it
with a 1MB size check. A lone record that still 413s is dropped, everything a transient
failure leaves undelivered goes back to the front of its queue, and the retry queue is
capped at max_queue_size so an unreachable workspace cannot grow memory without bound
* fix(azure_sentinel): retry undelivered records on the flush timer only
Requeued records made every later event cross the batch_size threshold, so a
down ingestion endpoint got one full-queue resend per request. Threshold sends
now go through flush_queue, so they take the flush lock instead of racing the
timer, and they stand down while records are awaiting retry.
A record that cannot be serialized raised out of the size probe and killed the
periodic flush task. The probe now runs inside the failure handling, so the
batch is split and only the record that cannot be serialized is dropped.
* fix(azure_sentinel): decide threshold sends under the flush lock
Concurrent callbacks all read logs_awaiting_retry before the first send
finished, so each one resent the whole queue once that send failed. The
flag and the batch_size threshold are now rechecked while holding the
flush lock, and each queue sends only itself instead of going through
flush_queue, which was retrying the other queue too.
* test(azure_sentinel): cover successful threshold waiters
* fix(azure_sentinel): preserve cancelled batches for retry
* fix(azure_sentinel): requeue only the undelivered part of a cancelled split
A batch over the ingestion cap goes out in pieces, so a cancellation partway
through requeued pieces the destination had already accepted and sent them a
second time on the next flush
The split helper now raises a cancellation carrying the records it never
delivered, and Azure Sentinel requeues those instead of the whole batch
* fix(azure_sentinel): drop batches a permanent rejection will never accept
A non-413 4xx from the ingestion endpoint or from the OAuth token call means the request
will fail the same way on every retry, so requeueing it held the batch, and every record
logged behind it, until the queue cap dropped them. Retryable statuses (5xx, 408, 429)
still keep the whole batch, and a shared classifier gives Datadog the same rule
The serialization probe now catches any exception, not just TypeError and ValueError,
because safe_dumps hands pydantic models to model_dump and can raise anything. It also
splits on record count, so a recovery flush sends batch_size records per request instead
of serializing the whole requeued queue to measure it
Both integrations re-raise a cancelled send as exactly asyncio.CancelledError. Python
3.12's asyncio.wait_for only translates the exact class into TimeoutError, so the
BatchSendCancelled subclass escaped the logging worker as an unhandled error
The awaiting-retry flag now follows the queue that survived the max_queue_size trim, so
a deployment with the cap at zero is not left waiting for a timer flush with nothing
queued to retry
* chore(logging): document mutable queue ownership
Annotate the queue detach and requeue constructions required by the logger's appendable queue contract so the type-discipline budget stays clean
* fix(datadog): preserve non-413 retry behavior
Keep Datadog's existing contract of requeuing every non-413 HTTP failure while Azure Sentinel applies its permanent-client-error policy through the shared splitter
* fix(batch_utils): requeue by default and let Sentinel opt into dropping
The shared splitter's default non-success handler is now requeue_after_http_error, the behavior Datadog had before the extraction, so a caller that omits the argument keeps its records. Azure Sentinel passes undelivered_after_http_error explicitly to drop permanent 4xx rejections
Also drops an explicit return None the strict ruff gate flags in the test helper
* fix(cloudzero): infer daily batch schema from every row
pl.DataFrame defaults to inferring column types from the first 100 rows,
so a day whose batch starts with more than 100 rows missing team_alias,
api_key_alias or user_email typed that column as Null and then raised a
ComputeError on the first row that had a value, failing the whole export
with a 500 and sending nothing.
Pass infer_schema_length=None when rebuilding each day's DataFrame, the
same guard the usage query already uses.
* test(cloudzero): cover late tag schema inference
Exercise the CloudZero resource tag field after a long run of missing values so a finite inference window fails the regression test.
* fix(cloudzero): preserve late resource tags
* style(cloudzero): remove redundant test comment
* fix(cloudzero): infer daily batch schema from every row
pl.DataFrame defaults to inferring column types from the first 100 rows,
so a day whose batch starts with more than 100 rows missing team_alias,
api_key_alias or user_email typed that column as Null and then raised a
ComputeError on the first row that had a value, failing the whole export
with a 500 and sending nothing.
Pass infer_schema_length=None when rebuilding each day's DataFrame, the
same guard the usage query already uses.
* test(cloudzero): cover late tag schema inference
Exercise the CloudZero resource tag field after a long run of missing values so a finite inference window fails the regression test.
A shadow eval job could only be scoped by identity, so "this user's traffic on model X
across every key they own" was not expressible and a models field on the start body was
silently dropped. The job now carries a models list that every target is narrowed to,
matched on the requested model group with model_group_alias resolved on both sides. An
unresolvable name is a 400 at start. Empty means every model, which is what every existing
row reads as. The dashboard start form gains an "Only on models" picker and the job
headline shows the scope.
* feat(otel): stamp litellm.request.route on the LLM call span
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(otel): drop redundant comment on REQUEST_ROUTE
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(otel): Final-annotate route test locals, drop field comment
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): read litellm.request.route off the server span
The LLM call span took the auth-normalized literal path from logging
metadata, which disagrees with the SERVER span wherever FastAPI matched a
template: on /engines/{model:path}/chat/completions the LLM span spelled the
model name while http.route carried the template, so the two spans grouped
into different buckets and the PR's premise did not hold.
Read the value off the span that already holds it. The request's root SERVER
span is anchored per request for parenting, and its attributes stay readable
after it ends, so request_root_http_route() answers from the async close
callback with the same http.route the SERVER span exports: the route template
on a normal route, the literal path where the passthrough hook rewrote it, and
the mount point on an MCP call. Nothing has to re-derive any of that, so the
two spans cannot drift apart.
The route the proxy recorded at auth stays as the backstop for a deployment
whose FastAPI instrumentation never mounted, where there is no server span to
disagree with. Off the proxy the attribute is omitted rather than empty.
---------
Co-authored-by: shivam <shivam@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Yucheng He <yucheng@berri.ai>
Suppression state moves out of request metadata into a request-scoped ContextVar.
refresh_proxy_server_request_body_snapshot copies metadata into
proxy_server_request.body, which deployments persist to spend logs, so the marker
naming each suppressed guardrail was readable by the caller whose request produced
it. Recovering it was enough to replay {token}:{name} for any CustomGuardrail and
switch off a PII or content-filter guardrail, since the check never verified the
named guardrail was a compression one. Nothing is read from metadata now, so there
is no marker to forge and the per-process token is no longer needed.
Routing-side compression reads the live messages instead of a pre-guardrail copy.
arm_pre_call runs before the pre-call hook, so its snapshot held the prompt as it
was before any masking guardrail rewrote it, and messages_for_routing handed that
to a compression guardrail which POSTs it to an external service. Masked content
left the proxy anyway. The cost is one combination: when the model hop compressed
and the hops differ, routing now classifies on the compressed text, since no
uncompressed copy survives that a masking guardrail has already seen.
policy_for_model no longer falls back to a marker scoped to tags the request does
not carry, which applied an 'eu' policy to a 'us' request on config order alone.
Each fix carries a regression test; all three fail when the fix is reverted.
* fix(shadow_eval): tell a tool-call shadow reply apart from an empty one
Both arrive at the attempt row as the same 'shadow router returned an empty
response', because _chat_final_text returns empty for a tool-final turn by
design and for a reply that genuinely carried no text. Those are different
things: an arm that chose a tool where the real model wrote prose is a
divergence a text judge cannot score, and the sampling side already drops the
real arm's tool-final turns for exactly that reason, so the shadow side reads
as a fault where the real side reads as a filter. A job that is almost all
'empty response' gives no way to tell a tool-happy arm from a broken one.
The error now names which of the two happened, and carries the finish_reason
and the routed model so the row says what the arm was doing. Every varying
part sits behind the first semicolon: operators read these by grouping on the
error text, and interpolating the model into the leading sentence would make
each row its own group.
The outcome stays 'error'. Whether a tool-call reply should instead be its own
non-judged outcome, excluded from the loss rate the way the real arm's
tool-final turns already are, needs the four aggregation predicates that spell
judged as outcome != 'error' rewritten, and a decision on how to surface the
new bucket. That is a separate change.
* fix(shadow_eval): read the tool name of a custom tool call
A custom tool call carries its name under custom.name with no function key,
so every one of them reported as tool=unnamed.
* feat(shadow_eval): judge tool calls instead of dropping the turn
A turn where either arm called a tool was discarded before it could be
compared: the real arm's at sampling, the shadow arm's as an error row. On
agentic traffic that is most of the traffic, so a job set to sample 10% was
sampling 10% of the prose-only slice. Tool calls now serialize to text on
every surface and are judged like any other response, and the judge is told
a tool call is not a defect so it scores the choice rather than the shape.
* feat(shadow_eval): show the judge what tools were available
Both arms were offered the same tools, but the judge only ever saw the
chosen call in isolation, with no way to tell whether a better tool existed
or the arguments matched what the tool expects. Threads the request's tool
definitions (name and description only) into the judge prompt, capped and
omitted entirely on turns that offered none.
* fix(shadow_eval): read a custom tool definition's name from custom, not function
A chat-completions custom tool definition nests name and description under
custom, mirroring how a custom tool call nests them (openai.types.chat.
ChatCompletionCustomToolParam). Reading only function rendered every one as
unnamed, telling the judge nothing about what it was.
* fix(datadog_llm_obs): keep the guardrail audit record under message redaction
Redaction nulled `guardrail_information` on the span whole, so an operator
running `turn_off_message_logging` (or a caller sending
`x-litellm-enable-message-redaction`) lost the record of which guardrails ran,
what they returned, and what they masked. Four of the record's fields can quote
the prompt; the rest report what the guardrail decided without reproducing it.
Replace only those four, the way
`_sanitize_guardrail_information_for_spend_logs` already does for spend logs,
and declare the field list once in `litellm/types/utils.py` so both readers
share it.
* fix(datadog_llm_obs): keep a lone guardrail record, and test through the span
Review round 1.
A guardrail that writes the metadata key itself leaves a single record where
the type says list, which Prometheus already normalizes at
`_guardrail_overhead_seconds`. Redaction dropped that shape and the latency
extraction raised on it, so the span was lost outright. Normalize once and use
it in both places.
The new tests now drive `create_llm_obs_payload` instead of reading the module's
private helpers and the record's declared field names.
The parser message alone cannot separate a judge that answered with nothing
from one truncated mid-object, and the two want opposite fixes. Records the
reply's shape, never its text, since no attempt row carries sampled content.
- Suppression markers now carry the per-process token `_pre_call_marker`
already uses, so a caller cannot switch off an always-on PII, content-filter
or compression guardrail by naming it in its own request metadata.
- Routing set to "none" with the model side compressed now classifies on the
pre-compression snapshot instead of the model-side guardrail's output.
- Both the proxy's pre-call arming and the router's routing hook resolve the
policy through one tag-aware `policy_for_model`, so an alias with several
tag-scoped markers can no longer suppress one marker's guardrail and then
route under another marker's policy.
- The pre-compression snapshot moved from request metadata to a ContextVar:
`refresh_proxy_server_request_body_snapshot` copies metadata into
`proxy_server_request.body`, which deployments persist, and the snapshot
holds the prompt as it was before any masking guardrail rewrote it.
- The compression selector lists Compresr guardrails too, not just Headroom.
An auto router marker deployment can now set auto_router_routing_compression
and auto_router_model_compression in its litellm_params, naming the
compression guardrail each hop should use (or "none" for no compression on
that hop). Neither key set means the request's own compression guardrails
keep applying to both hops unchanged.
Backend: Router.async_pre_routing_hook resolves the marker's policy and
compresses a copy of the messages for the routing decision only when the
policy differs from what the model call already got; when both hops share
the same compression, it reuses what the ordinary pre-call guardrail
pipeline already produced instead of compressing twice. The proxy layer
suppresses every other compression guardrail once a policy is engaged and
arms the model-side guardrail even when it is not default_on.
UI: the auto router's Detailed Configuration gains an Advanced: Compression
section with a routing-decision selector and a same/different toggle for
the model call, matching the same/different address pattern.
The prior commit claimed claude-sonnet-5 reasons invisibly by default and eats
the judge's budget regardless of what the call asks for. Verified against a
live proxy: with no thinking param (what _call_judge sends today), forced
tool-choice json_mode, native structured output, and even an explicit
thinking=adaptive, the model returned 0 reasoning tokens and a clean compact
verdict every time, on prompts up to several thousand characters.
The real mechanism only shows up with an elevated reasoning_effort or
output_config.effort on the request, which happens when the judge_model
deployment is configured with one, e.g. an admin pointing the judge at their
best reasoning model. Reproduced directly: reasoning_effort=max, 300-token
cap, real Anthropic reply came back finish_reason=length, content=None, 299
of 300 tokens spent on reasoning. Same request at 4096 returned a valid
verdict. This is a narrower, verified claim than the one it replaces.
The cap covers reasoning tokens as well as the verdict, and the models people
pick as judges reason before answering whether the call asks them to or not:
Anthropic's 5 family thinks adaptively and cannot be told not to, so the
reasoning bills against max_tokens with nothing in the request to opt out.
At 1500 the reasoning consumed the budget and the reply arrived empty or cut
off mid-object, which the attempt recorded as an unparseable judge verdict
rather than a result. Headroom costs nothing: max_tokens is a ceiling and only
generated tokens bill, so the only movement is that judge calls which used to
bill their full budget and return nothing now return a verdict.
Deliberately not passing reasoning_effort to bound the reasoning instead:
is_thinking_enabled treats any reasoning_effort as thinking-enabled, which
drops the forced tool_choice that json_mode relies on and turns thinking on
with a 1024-token floor for judges that were not reasoning at all.
* fix(azure): restrict the storage credential chain to deployment identities
The keyless Azure Storage path walks the full DefaultAzureCredential chain, so a
proxy with no storage service principal authenticates as whichever identity the
host happens to carry: an operator's az login on a workstation, or the
AZURE_CLIENT_ID/AZURE_CLIENT_SECRET service principal set for Azure OpenAI.
Neither is the identity granted Storage Blob Data Contributor.
Narrow the chain to workload identity and managed identity, the two credentials
a deployment legitimately holds. Azure OpenAI, Postgres IAM auth and the other
callers of get_azure_ad_token_provider keep the full chain.
* test(azure): read the credential chain off the mock instead of an accumulator
* chore: drop a stray launch traceback committed at the repo root
* fix(azure): let the storage chain reach a system assigned managed identity
DefaultAzureCredential keeps one managed identity link and pins it to
AZURE_CLIENT_ID, so a host that sets that variable for Azure OpenAI and runs as
a system assigned identity never got asked for a storage token. Build the chain
from the three credentials a deployment can carry instead of subtracting the
ones it cannot.
One unreachable vector store used to wipe out every store's context on a
chat completion carrying vector_store_ids: the search raised, the blanket
handler returned the original messages, and the request answered with no
retrieved context at all. Each store's search now has its own handler that
warns with the vector store id and moves on to the next store.
The same loop appended every store's results to the original messages
instead of the running copy, so with two healthy stores only the last one
reached the model. It now chains through modified_messages.
The Router is injected through a ProxyRuntime protocol instead of an
in-function litellm.proxy.proxy_server import, so the hook's routing can
be driven in tests without touching proxy globals.
* feat(datadog_llm_obs): cost tag dimensions, router decision fields, reasoning token metric, redaction gating
* test(datadog_llm_obs): satisfy test quality gate
* fix: forward integer parent_id as its string form
* fix(datadog): sanitize redacted message roles
* fix(datadog): keep the A2A agent role on redacted spans
* fix(datadog): merge current staging budget
* style(datadog): format redaction tests
* fix(datadog): handle malformed redacted roles
* test(datadog): put the test quality suppression on the reported line
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>
* feat(azure): support credential chain for storage
* test(azure): clarify credential seam suppressions
* fix(azure): read chain tokens in a worker thread
The credential chain walk (IMDS probe, CLI subprocess) is blocking I/O,
so reading the provider inline in async set_valid_azure_ad_token stalls
every request on the worker's event loop
* fix(guardrails): run apply_guardrail-only providers in logging_only mode
A CustomGuardrail that implements only apply_guardrail inherited the CustomLogger
no-op async_logging_hook, so mode: logging_only never scanned anything and never
recorded guardrail_information. CustomGuardrail.async_logging_hook now routes the
logged request and response through the call type's guardrail translation on
copies and appends the verdict to standard_logging_object.guardrail_information.
Resolves LIT-4876
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): keep logging_only scan copies inside the error boundary and return a fresh logging payload
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(guardrails): cover embedding scan, native-hook bypass, and unmapped call type in logging_only
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The LLM Obs callback copied litellm's OpenAI-shaped objects into the span
verbatim, so every field Datadog names differently landed somewhere it does
not read: tool calls kept their nested `function` wrapper instead of DD's
name/arguments/tool_id, tool messages carried no result linking them to their
call, the request's tools were never sent, and prompt-cache counts sat inside
meta.metadata rather than the span metrics its cache dashboards chart.
One rule governs the message mapper: add the fields Datadog declares, and never
destroy content it did not understand. Content collapses to its text only when
it has text, so a content list carrying tool or image blocks rides along
unchanged, and absent messages map to an empty input rather than a fabricated
turn. Tool calls and results are read from both dialects, the OpenAI
`tool_calls` / `role: tool` shape and the Anthropic `tool_use` / `tool_result`
content blocks, so /v1/messages sessions gain tool linking they never had.
Cache counts come from the same owners the savings dashboard uses, so every
provider spelling resolves through one place rather than a second local guess.
The three cache metrics partition the input count: litellm's normalized prompt
total includes both cache categories, as the cost calculator's pricing helper
documents, so the non-cached residual subtracts reads AND writes. Counting a
primed prefix as ordinary input had inflated non-cached usage by exactly the
cache-write count on every priming request.
Correlating a result to its call reads ids and names structurally and parses no
arguments, so a tool call's arguments are decoded once per span rather than
once per pass, and arguments past a size bound ship as the raw string instead
of paying a decode that multiplies memory on hostile compact JSON.
The flat `output_tool_calls.*` metadata copies go away with this: they were a
second representation of a fact that now has its own field on the same span.
The caching-savings marker litellm_gateway_injected_cache credits gateway-earned
prompt-caching savings to the deployment it names, or to every deployment via
the empty-string sentinel. Two paths lost that scope:
- the router prompt-management factory stamps a provisional deployment's
model_info into kwargs before the prompt pass runs, so an injection recorded
there named that provisional pick and a differently-billed deployment lost
the credit
- record_gateway_injection overwrote on every positive delta, so a per-leg
stamp (the Bedrock converse tool_config one included) downgraded an
existing every-deployment mark and the leg billed after a failover lost
the credit
record_gateway_injection now takes injected_for_every_deployment, the two
pre-choice callers declare it, and an every-deployment mark is never narrowed
by a later per-leg stamp. Per-leg marks still overwrite each other. Spend
amounts are untouched; only the savings attribution is affected.
Also unblocks make lint at the staging tip: tests/e2e/test_junit_properties.py
landed three basedpyright reds via an e2e-only PR whose lint job skipped, now
suppressed as the deliberate duck-typed double they are.
* feat(alerting): slack alerts for per-user daily/monthly spend thresholds and spend anomaly detection
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(alerting): use specific ValidationError matches in config rejection test
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): tolerate mocked slack alerting args when scheduling user spend scan
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(alerting): reject non-finite values in user spend alert settings
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
