The accumulated-JSON fallback ran json.loads over the whole buffer after every fragment and, on failure, kept the buffer without resetting it. A buffer that ever held more than one concatenated JSON value could never parse (json raises on trailing data), so it returned None on every subsequent chunk while growing without bound - an unrecoverable per-request CPU spin. Parse one value at a time from the front with raw_decode and keep the remainder, draining trailing values on later calls and at end of stream.
The async streaming error paths fired the sync failure_handler in a thread
and the async_failure_handler via create_task at the same time, so both
mutated the shared logging object concurrently and could crash pydantic-core.
Route failure logging through a single guarded dispatch_failure_handlers, so
the sync handler only runs after the async one completes.
Extract the inline sync streaming post/decode into make_sync_call so it can be
exercised with an injected client, mirroring make_async_call, and add a sync
regression test that each token is forwarded after exactly one pulled frame.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Mirror the sagemaker_chat fix on the native sagemaker/ streaming path: the
sync and async completion handlers read the invocations-response-stream body
with iter_bytes(chunk_size=1024) / aiter_bytes(chunk_size=1024), so httpx
withholds bytes until 1024 accumulate and tokens arrive in gap-then-burst
waves. Drop the fixed chunk size so each decoded event is forwarded as its
bytes arrive.
Also add a boundary-agnostic decoder test proving frames reassemble correctly
regardless of where transport reads split the stream.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The huggingface embedding test fixture reloaded
litellm.llms.custom_httpx.http_handler, creating a new HTTPHandler class
object. llm_http_handler keeps the class captured at import time, so any
test running later in the same process that injects a client built from
the reloaded class fails the isinstance check and the mock is silently
discarded, causing a real network call. Under pytest-xdist loadscope this
surfaced as a deterministic failure of
test_accept_header_in_completion_request_jwt whenever an unrelated PR
shifted worker distribution.
Also removes the same reload pattern from the vertex rerank integration
test (both were previously removed in a6df01caec and resurrected by a
merge conflict resolution) and hardens the agentcore victim test by
dropping the bare except that swallowed the real error
* fix(anthropic): strip uniqueItems + other unsupported array/object constraints from output_format schema
Anthropic's structured outputs (`output_format`) validate the JSON schema
against a strict subset and reject cross-element / count constraints that a
constrained-decoding grammar cannot enforce, returning a 400
`invalid_request_error`.
`filter_anthropic_output_schema` already stripped the numeric / string /
item-count constraints (minimum, maximum, exclusiveMinimum/Maximum, minLength,
maxLength, minItems, maxItems) but still let these through:
- uniqueItems
- contains / minContains / maxContains
- minProperties / maxProperties
so a request using them fails with e.g. "output_format.schema: For 'array'
type, property 'uniqueItems' is not supported".
This is provider-visible: newer Claude models on the native `output_format`
path (e.g. `azure_ai`) 400, while `vertex_ai` is unaffected because it is
forced onto the permissive tool-use path (#18625 / #19201).
Add the missing keywords to the unsupported-field set and the description map,
and skip the advisory description note for a disabled boolean constraint
(`uniqueItems: false`) so it isn't misdescribed as required.
* fix(anthropic): serialize contains sub-schema in output_format advisory note
Address Greptile review: the `contains` advisory note previously discarded the
sub-schema, so the description only said an item must match "a schema" without
saying which. It now serializes the sub-schema as JSON (e.g. "array must
contain an item matching: {\"type\": \"integer\", \"const\": 1}"), matching the
other stripped constraints which carry their value. Sub-schema (dict/list)
values are json.dumps'd; scalar constraints are unchanged.
* style(anthropic): apply ruff format to output_format filter change
* style(test): ruff format anthropic schema filter tests
* test(anthropic): cover output_format array/object constraint filtering in test_litellm tree
Mirrors the schema-filter tests under tests/test_litellm/ so the coverage
job exercises the new uniqueItems/contains/min-maxProperties handling and the
uniqueItems: false branch.
---------
Co-authored-by: Darien Kindlund <darien@kindlund.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(guardrails): add only_scan_new_messages for per-session incremental scanning
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* fix(guardrails): use fixed TTL constant and revert unrelated test formatting
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* fix(guardrails): run only_scan_new_messages in the unified apply_guardrail path
The initial wiring lived in BedrockGuardrail.async_pre_call_hook, but the proxy
routes Bedrock through the unified apply_guardrail interface, so the flag had no
effect live. Move incremental selection into apply_guardrail: filter the flat
texts list against per-session scanned hashes, skip the Bedrock call when nothing
is new, and mark hashes only after a successful (non-blocked) scan. Full-context
fallback is preserved when there is no session id, the cache is unavailable, or a
masking guardrail is configured.
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* test(guardrails): cover session-id fallbacks and mark_texts_scanned guards
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* fix(guardrails): fall back to full scan when incremental guardrail masks content, use shared cache
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* test(guardrails): cover generic agent multi-turn incremental scan
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* test(guardrails): cover incremental scan cache resolver fallbacks
Co-Authored-By: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
* test(guardrails): cover flag interactions and /v1/messages incremental scan semantics
* feat(guardrails): make GUARDRAIL_SCANNED_MESSAGES_CACHE_TTL_SECONDS env configurable
* test(guardrails): prove skip_system/skip_tool are enforced upstream of incremental scan
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Ishaan Jaffer <155045088+ishaan-berri@users.noreply.github.com>
Co-authored-by: Yucheng Zhu <yucheng@berri.ai>
Azure AI Foundry and Vertex AI serve Claude on the first-party Anthropic
Messages contract, which was verified live to be byte-identical to
api.anthropic.com: a leading role:"system" entry in messages is rejected on
every model ("messages.0: use the top-level 'system' parameter"), and a
mid-conversation role:"system" reminder is accepted in place on Claude 4.8+/5
but 400s on Claude 4.7 and older ("role 'system' is not supported on this
model"). This is the same contract Bedrock Invoke already handles model-aware
(PRs #32578/#32831/#32882); Vertex and Azure did no hoisting at all, so a Claude
Code session on an older Vertex/Azure Claude model hard-400s on its reminder
turns, and the only thing sparing 4.8+/5 was that nothing was hoisted
Extract Bedrock's model-gated normalization into the shared
AnthropicMessagesConfig base as _normalize_system_role_messages and call it from
the Vertex and Azure messages configs. Flagged models (4.8+/5) hoist only the
leading run of system entries and keep mid-conversation reminders in place so
the top-level system prefix stays byte-identical and the prompt cache is
preserved; unflagged models hoist every system entry so the request returns a
completion instead of a 400
Add supports_mid_conversation_system to the azure_ai and vertex_ai Claude 4.8+/5
cost-map entries. Exact cost-map hits win over the claude-mid-conversation-system
fallback rule, so without the explicit flag those models would be treated as
unsupported and hoist every reminder, collapsing the prompt cache (the exact
customer regression). A per-provider test guards this so future 4.8+/5 entries
cannot silently miss the flag
Closes the Vertex/Azure gap from the customer RCA
Caller-supplied bedrock_tags land as AWS resource tags under the proxy's
AWS identity, letting an authenticated caller forge ownership or
cost-allocation labels. Add bedrock_tags to _BANNED_REQUEST_BODY_PARAMS
so per-request tags need general_settings.allow_client_side_credentials
or configurable_clientside_auth_params on the deployment, matching the
aws_bedrock_project_id precedent. Deployment-level bedrock_tags in
litellm_params are unaffected.
Also stop an explicit empty bedrock_tags list in litellm_params from
falling through to optional_params
* fix(fireworks_ai): set Content-Type application/json in validate_environment
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(fireworks_ai): delegate chat validate_environment to OpenAIGPTConfig
Instead of re-adding the JSON Content-Type default inside FireworksAIMixin,
FireworksAIConfig now delegates header construction to OpenAIGPTConfig and only
layers the Fireworks-specific x-session-affinity header on top, so the
Content-Type default can no longer drift away from the OpenAI base and reintroduce
the 415.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(fireworks_ai): cover missing api key error path in chat validate_environment
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Krrish Dholakia <krrishdholakia@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
glm-5p2 (and its fireworks_ai/glm-5p2 alias) carried cache_read_input_token_cost
of 2.6e-07, the GLM 5.1 rate; the entry was seeded from the wrong row. Fireworks'
standard serverless rate for GLM 5.2 is $0.14/1M = 1.4e-07, so every prompt-cache
hit was billed at nearly double the real rate.
Corrects the value in both the canonical map and the bundled backup. The existing
fireworks cost-calculator test now reads the cached rate from the map instead of
hardcoding it, so it tracks the shipped value.
The /v1/messages handler resolved only the auth object and the trace id, so tool
listing and tool execution ran without the caller's MCP auth headers. That fails
quietly rather than loudly: the tool still executes, just with no credentials, so
every server behind interactive OAuth, a bearer token or per-user env vars returns
nothing while the model reports it has no access. Only a no-auth server looks
healthy, which is exactly what the first proof used.
Threading the missing arguments would have left the real problem in place. Each
gateway surface rebuilds the same context by hand (responses/main.py twice,
chat_completions_handler, mcp_streaming_iterator), which is why a new surface
drops fields; this adds a fifth that dropped six of eight. Resolve it once into a
frozen MCPRequestContext and have the handlers take that, so a field cannot be
forgotten at a call site. chat_completions_handler now uses it too, and the
resolver reads user_api_key_auth from both metadata keys because
LITELLM_METADATA_ROUTES carry it in litellm_metadata while chat uses metadata.
Also stop the loop when every tool call was skipped. tool_results is empty then,
and the tool_result message built from it has empty content, which Anthropic
rejects; the caller saw a 400 from mid-loop instead of the model's own answer.
Tests pin both: dropping the headers from either listing or execution fails, and
so does removing the empty-results guard.
MCP tool calling worked on /v1/chat/completions and /v1/responses but not on
/v1/messages. Those are the only two surfaces with an MCP gateway entry point,
so a litellm_proxy MCP reference reached Anthropic verbatim inside tools and the
API rejected the request with "Input tag 'mcp' found using 'type' does not match
any of the expected tags". The playground never surfaced this because it dropped
the reference before sending, and disabled the MCP selector for the endpoint.
Add the third entry point in anthropic_messages_handler, ahead of the provider
branch so it covers the native path and both bridges from one place. The gateway
expands the reference against the caller's own credentials and access control,
which is the whole point of routing it through litellm rather than handing the
url to the provider.
/v1/messages needs Anthropic's own tool shape, so transform_mcp_tool_to_anthropic_tool
joins the OpenAI chat and Responses transforms alongside it. The tool loop speaks
tool_use and tool_result rather than OpenAI tool_calls, and reuses the existing
FakeAnthropicMessagesStreamIterator to re-stream the result, the same pattern the
websearch interception already uses on this route. Argument extraction moves into
the shared extractor: an Anthropic tool_use block carries its arguments under
`input`, and reading only `arguments` failed silently, executing the tool with
every argument dropped.
On the frontend the request builder declared selectedMCPTools and never read it,
so no tools key was ever sent. Wire it through a shared block builder and add the
endpoint to MCP_SUPPORTED_ENDPOINTS, which is what greys the selector out.
Resolves LIT-4517
Resolves LIT-4518
* feat(guardrails): add Compresr guardrail for query-aware context compression
Adds a first-class guardrail that compresses bulky message content (tool
outputs, RAG chunks, search results) through the Compresr API before the
request reaches the LLM, via the apply_guardrail / structured_messages hook
so it covers /chat/completions, /v1/messages, and /v1/responses (the latter
through the texts channel, mirrored only when the replacement is
unambiguous; anything ambiguous is left uncompressed).
Distinct from whole-conversation compressors:
- Query-aware: each message is compressed against the intent that produced
it (a tool output against its originating tool call's name + arguments,
resolved via tool_call_id; otherwise the last user message).
- Recoverable: each compressed message carries a hash marker and the request
gains a compresr_retrieve tool, so the model can pull the original content
back through the agentic loop when the compressed version is not enough.
Originals are cached in-process, scoped to the caller's virtual-key hash
plus the request's litellm_call_id, with a TTL and a per-call byte cap;
recovery is skipped when no caller scope is available so one caller can
never read another's originals. The store is per-process, so multi-worker
deployments need sticky routing (or enable_retrieval=false).
Fail-closed by default (fail_open configurable), SSRF-validated api_base
(alternate IP-literal encodings included), cross-tenant-isolated recovery
store, and upstream errors redacted from client-facing responses. The
outbound client follows redirects and re-resolves DNS per request, so the
api_base host/IP checks are defense-in-depth, not a full SSRF guarantee;
this is documented as a known limitation. Requests where nothing was
actually compressed are returned untouched (same object identity) so
handlers skip the write-back. Auto-discovered via the guardrail_hooks
registry.
* fix(guardrails): cap Compresr recovery store total memory
The recovery store bounded bytes per call and entry count, but had no
aggregate cap: 256 tracked call ids at the 10 MiB per-call default could
retain ~2.5 GiB per worker. A flood of requests with distinct
x-litellm-call-id values and large compressible tool outputs could
exhaust a shared proxy worker.
Add a global byte budget (_MAX_TOTAL_STORE_BYTES, 256 MiB) across all
entries. A running total is maintained on every insert/eviction so the
cap is enforced without re-encoding the whole store on the request path;
oldest entries are evicted once the budget is exceeded, always keeping
the most-recent entry so recovery still works for the request populating
the store. +2 regression tests.
* fix(guardrails): gate and bound Compresr recovery loop
Two hardening fixes to the compresr_retrieve agentic loop:
1. Only run the loop when a retrieve call resolves to recovery state this
guardrail actually created for the request. Previously the gate checked
only that the caller-supplied tool list contained a compresr_retrieve
function and that the model emitted a call, so a caller could define
their own same-named tool and force an extra provider round-trip with
nothing to recover. The plan now returns run_agentic_loop=False when no
requested hash resolves.
2. Bound the follow-up against retrieval amplification: each distinct hash
is expanded at most once (repeats get a short marker) and at most
_MAX_RETRIEVALS_PER_LOOP calls are honored, so prompting the model to
call compresr_retrieve many times with the same marker cannot balloon
the follow-up. _retrieve_original now returns None on miss.
+3 regression tests; two existing security tests updated to assert the
stronger veto behavior (forged/cross-tenant hashes now stop the loop
entirely instead of returning a not-found follow-up).
* fix(guardrails): warn when Compresr recovery is skipped without auth scope
When enable_retrieval is on (the default) but the proxy has no per-key
auth, the request has no caller scope, so recovery is silently disabled:
content is compressed but the compresr_retrieve tool is never injected and
the originals are dropped, with no runtime indication. Emit a one-shot
call-time warning so operators can see recovery is being suppressed and
configure virtual-key auth. +1 regression test.
* style(guardrails): tighten Compresr guardrail comments
Condense the verbose multi-line inline comments and the api_base docstring
to concise form. No behavior change.
* fix(guardrails): keep injected tool on Responses API + bound recovery markers by byte cap
Two fixes for reviewer-flagged defects in the Compresr guardrail:
- Responses API: _merge_tools_after_guardrail iterated only over the
request's original tools, dropping any tool a guardrail appended (the
compresr_retrieve recovery tool) whenever the request already had tools.
Keep the appended tools so recovery works on /v1/responses.
- Recovery markers: markers + originals were built for every compressed
target before the per-call byte cap trimmed the store, so an evicted
original left a marker the model could never retrieve. Attach recovery
only while the store (existing entries under the same key + this call's
originals) stays within the cap, so a shipped marker is always retrievable
-- including on a later turn that reuses the store key.
Adds regression tests for both paths.
* refactor(guardrails): extract _existing_originals to keep apply_guardrail under the complexity gate
The byte-cap fix added a branch to apply_guardrail, tipping it past the
C901 complexity ceiling. Move the store lookup into a small helper; no
behavior change.
* fix(guardrails): harden Compresr SSRF blocklist, re-arm no-scope warning, tolerate odd tool shapes
* fix(guardrails): rerun input guardrails on Compresr retrieval follow-up
* chore: remove unrelated deepkeep files committed by mistake
---------
Co-authored-by: charafkamel <charafkamel@live.com>
