Buffered streams governed by post_call policy pipelines now deliver text
rewrites back into the stream per surface (chat SSE, responses SSE,
anthropic messages SSE) instead of rejecting the request with a 400
upfront. Rewrites chain across pipeline steps; tool-call rewrites and
translations without stream write-back still withhold the stream.
- forward unrouted /gigachat/* requests with env credentials like other passthrough providers (the old fallback returned 400 on any request without a routed model, /gigachat/models included)
- fix basedpyright budget breaches across the gigachat provider, common_request_processing, and llm_passthrough_endpoints with real narrowing, no new suppressions
- add regression tests for the fallback target, auth header, and model-less endpoints
_input_item_provenance converted every input prefix, so an n-item request paid
for n+1 full conversions. It now converts each item once, glues consecutive
function_call items (plus their trailing-assistant context) into units so the
transform's tool_call merging is reproduced inside the unit conversion, and
verifies the unit concatenation against one full conversion, bailing to the
full-conversion fallback on any mismatch. Messages from multi-item units are
tainted, which keeps parallel tool calls patchable exactly like the old prefix
pass while unpredicted merges fall back safely.
A guardrail handing back a non-list structured_messages payload (the
HiddenLayer v2 evaluation dict) previously fell through the length-mismatch
fallback and 500ed converting the dict's keys as messages. The write-back is
now skipped for non-list payloads, restoring the previous no-write-back
behavior on the Responses surface.
Also refreshes the compresr texts-mirror docstring, which still claimed the
Responses translation cannot round-trip structured_messages.
The route-level regression test returns a real prisma row from a mocked
update and asserts both routes serialize it to a 200 with the toggled
blocked flag, which is exactly the path that raised AttributeError before
the validator guard. Also binds the loop variable in the e2e poll lambda
(ruff B023).
API_ROUTE_TO_CALL_TYPES listed the sync llm_passthrough_route first, so every
call_types[0] consumer resolved /llm_passthrough to a call type with no
guardrail translation handler, and the {model}:generateContent patterns never
matched a concrete route because the placeholder segment carries a literal
suffix the matcher treated as an exact segment. Reorder the passthrough
entries async-first, teach the matcher placeholder-with-suffix segments plus
suffixed multi-segment tails (mirroring FastAPI's {model_name:path}), add the
missing /v1beta generateContent entries, and register a Google GenAI
guardrail translation handler so guardrails actually scan generateContent
requests, responses, and streams.
Azure's chat completions validator rejects tool parameters carrying a
top-level anyOf/oneOf/allOf for every model family. AzureOpenAIConfig and
the o-series config now flatten them via the shared helper moved to
prompt_templates common_utils. Requests bridged to the Responses API for
gpt-5.4+ with reasoning active keep the union, which that surface accepts
The sync __next__ exhaustion branch stores calculate_total_usage() in the
final chunk's _hidden_params when stream_options is None, but the async
__anext__ sibling branch never did. Converted (fake) streams, like the ones
the Headroom guardrail produces by flipping streaming /v1/responses calls to
non-streaming, are consumed async, so their real usage never reached the
completion-to-responses bridge and it token-counted from scratch, reporting
input_tokens=0. Mirror the sync branch's hidden-usage block into the async
exhaustion branch and add a regression test that async-iterates a
CustomStreamWrapper over a MockResponseIterator and asserts the final chunk
carries the mock response's usage.
* fix(otel/v2): detach credential-routed tenant spans into their own trace
Multi-tenant OTel v2 routes a team or key's LLM-call span to that tenant's
own vendor account (New Relic, Arize, Langfuse, Weave) via dynamic OTLP
credential headers, while the request-root, auth, and db spans stay on the
operator's default backend. The span was still parented into the request
trace, so the tenant account received a child whose parent it never got,
and New Relic rendered it as a fragmented trace with a missing parent.
Detach a credential-routed span the same way a project-routed (Phoenix)
span already detaches: root a fresh trace in the tenant account and link
back to the request trace for correlation. Service-name routing keeps
parenting, since it only relabels service.name on the same operator
backend where the parent is present.
Guard the detach on the callback actually owning an OTLP exporter the
credentials can reach: a callback owning only a console or in_memory
exporter has nowhere to stamp them, so the span would export to the
default backend unchanged and detaching would orphan it on the very
backend that holds its parent. In that case warn once and keep the
default tracer.
* fix(otel/v2): derive tenant-route routability from resolved exporter transport
A denylist classified an owned exporter as routable whenever its kind was
not console/in_memory, so a typo'd or unavailable kind (e.g. "otlp",
"grcp") passed the check while _exporter_from_spec falls it back to a
header-ignoring console exporter. Detaching such a span would root a fresh
trace that only ever reaches the operator console, never the tenant
backend, orphaning it on both sides.
Route on a shared exporter_transport() predicate that resolves the kind the
same way _exporter_from_spec builds it (registered factories + otlp_http
aliases -> http, otlp_grpc aliases -> grpc, else headerless), so an
unresolvable kind is headerless and stays parented. Fixes the same latent
gap in project routability.
The one-time seed for a budget window row subtracted the batch's own
LiteLLM_SpendLogs rows by request_id, and request_id is the client's
x-litellm-call-id whenever the response carries no id of its own. Carrying
that set through the queue meant an unbounded, client-controlled aggregate
that the commit-failure requeue kept alive across retries.
Every log row at or after a batch's earliest start is owed by an increment
that still reaches the row, so summing only rows before it needs nothing
from the request. That drops request_ids end to end and closes the
cross-pod double count the id list could not see.
The driver retries rate-limit-shaped failures twice with a 65s backoff, but
RATE_LIMIT_SHAPED_RE only matches 429 / rate limit / too many requests /
throttled / CLI timeout. A saturated provider does not always say any of
those: litellm-e2e-pr build 182 turned a green cell red on
status 503: litellm.ServiceUnavailableError: BedrockException -
{"message":"Bedrock is unable to process your request."}
which matched nothing and so was never retried.
Add TRANSIENT_UPSTREAM_SHAPED_RE for 503, 529, "service unavailable",
"overloaded" and Bedrock's "unable to process your request", and have the
retry loop ask is_retryable_shaped (either shape) instead of
is_rate_limit_shaped.
Kept as a second pattern rather than widened into the first on purpose. The
conftest feeds RATE_LIMIT_SHAPED_RE into the rate-limit summary, which the
binary-search helper reads to decide whether to lower a provider's request
rate. A 503 says the provider is out of capacity, not that we are asking too
often, so folding it in there would keep ratcheting our rate down against a
condition our rate never caused.
Retry-loop logging now names which shape fired, so a saturated upstream is
distinguishable from a throttled one in the run output.
Verified against the literal failure text from build 182: is_retryable_shaped
is True while is_rate_limit_shaped stays False.
Note this covers CLI-driven rows only. HTTP-probe rows (tool_search,
count_tokens) have no retry layer at all, so the specific cell that failed in
build 182 is still unprotected -- that is a separate change to http_probe.
#34940 widened the mask-in-place safety guard so a Responses-API
`instructions` field (and a combined messages+input body) skips the
PII masking branch. With `on_flagged: "monitor"` that fell straight
through to "allow", so PII that used to be masked now reaches the
model unredacted.
Monitor means "don't block", not "don't redact". Recover the one shape
whose payload is still fully writable: mask it and write the redacted
instructions back into `data["instructions"]` directly, since
apply_redacted_messages_back has no path for that field and would
otherwise fold the instructions text into `data["input"]`.
The combined messages+input and multimodal shapes stay unmasked - both
are unsafe to write back, not merely unwritable - and now log an error
naming the reason instead of passing silently.
No block/allow decision changes: block and inject_system_message keep
the exact outcomes #34940 shipped.
Direct access on /model/info was read from the user record alone, so an
unrestricted user calling with a key limited to a few models saw every non-team
deployment, including ones the key gets a 403 on. Resolve the key's grant the
same way and intersect the two.
Resolving a grant now also expands access groups, which the key path needs and
the user path was missing.
Claude-Session: https://claude.ai/code/session_01XL7LBFEew4wi8gphVCDq6n
