A terminal HTTP failure partway through chunking now logs the summed usage
and cost of the ApplyGuardrail calls AWS already billed, mirroring the
blocked-chunk path.
A blocked chunk now logs the summed usage and cost of every ApplyGuardrail
call AWS billed for the logical request, not just the blocking chunk.
Client-supplied metadata.standard_logging_guardrail_information is stripped
at the proxy boundary so callers cannot forge (even negative) guardrail
cost into spend, and guardrail_information_cost ignores negative or
non-finite entry costs as defense in depth. The cost map schema test now
allows guardrail_cost_per_unit and the guardrail mode.
A model with a long time-to-first-token leaves the proxy's response completely
idle, so any hop with an idle read timeout (AWS ALB and nginx both default to
60s) drops a connection that is perfectly healthy and would have delivered its
tokens shortly after.
The keepalive engines LiteLLM already ships wrap the response object, so they
fill a gap once the upstream has answered and then gone quiet. They cannot fill
the gap before it answers at all, and that is where the whole wait is spent:
measured against api.openai.com/v1/chat/completions with gpt-5.6 at
reasoning_effort high, the response headers and the first body byte both arrive
at 37.90s. Nothing has entered the ASGI response phase by then.
The upstream call is now raced against the keepalive interval, and when it
loses, the SSE response is opened immediately and ": ping" comments, which every
conformant SSE client ignores, fill the wire until the real response is ready to
be replayed onto it. One seam per funnel: base_process_llm_request covers every
native route, create_pass_through_route covers every passthrough route.
Committing the status line that early is the cost. A failure discovered after
the first ping reaches the client as an SSE error frame under a 200 rather than
as an HTTP error status, and LiteLLM's own x-litellm-* response headers are not
yet known. keepalive_ping_has_fired already documents the same trade-off for the
existing engines. Both are why this stays off until an operator sets
litellm_settings.sse_keepalive_ping_interval_seconds.
Separately, the passthrough relay reached neither engine even for mid-stream
gaps, which is the shape of #32491 and #24929, so the relayed bytes get the same
treatment, gated on the upstream declaring text/event-stream and only emitted
between complete frames so a binary transport (AWS event streams on /bedrock)
and a stall halfway through a frame are both left alone.
Fixes#34819
Price ApplyGuardrail usage units recorded by PR #37225 with a new
bedrock/guardrails entry in the model cost map (regional override via
bedrock/{region}/guardrails), add the per-request guardrail_cost to the
standard logging payload's response_cost and CostBreakdown, surface it in
the x-litellm-response-cost header, and bill blocked requests through the
failure hook so key and team budgets see what AWS bills
GET /guardrails/usage/overview and GET /guardrails/usage/detail/{id} 500ed on a
database that has not applied 20260817143646_add_daily_guardrail_usage_units yet
(pip installs on litellm-proxy-extras 0.4.86 with DISABLE_SCHEMA_UPDATE=true).
Both endpoints now return their metrics with empty units and log one warning
until the migration lands.
The daily guardrail metrics and usage-unit upserts are non-idempotent
increments, but the retry loop re-sent every failed row on any exception.
An ambiguous post-send failure such as a read timeout after the write had
already committed therefore stacked a second increment and inflated the
billable unit totals served by the guardrail usage endpoints.
Retry only DB_RETRY_SAFE_ERROR_TYPES (httpx.ConnectError), the same rule
the spend writer and autorouter rollup use for increment upserts, and log
any other failure once as terminal for that row while the rest of the
batch still lands.
Follows up #37225
A transient DB error during the spend log flush dropped that batch's guardrail
metrics and usage unit rows for good. Retry only the rows that failed, up to 3
times with 1s/2s/4s backoff, mirroring the daily spend writer, and inject the
sleep so tests stay fast. Lowers the lint budgets the refactor freed up
The flush and the usage endpoints summed units with a scan per distinct key,
quadratic in rows times keys; group sorted rows instead. Skip payloads without
a request_id like the metrics path, type the flush key as a NamedTuple, and drop
the (guardrail_id, date) index that the primary key already covers
Per-row guards in the daily metrics and usage unit flush so a single DB error no longer drops the rest of the batch, plus removal of narrating comments flagged in review
The Noma guardrail sends the conversation to the scanner in `inputs`. It
also forwarded `request_data` whole, which repeats that same conversation
under `messages` (or `input` on the responses API), and attached
`logging_obj.model_call_details`, which repeats it a third time.
For image-heavy calls that duplication is most of the request. A
production scan of a request carrying base64 images measured 100MB total,
of which 94.8MB was `request_data` against 5.1MB of `inputs` - the proxy
was uploading ~95% redundant bytes, and paying to serialize them.
Drop `messages` and `input` from `request_data` and from
`model_call_details`. This is a denylist rather than an allowlist on
purpose: every other key is still forwarded untouched, so a scanner-side
change that starts reading a new `request_data` key needs no matching
release of this hook. The removed keys are ones the scanner never reads -
it takes context only from metadata, litellm_metadata,
provider_specific_header, litellm_session_id/trace_id/call_id, stream,
response/responses ids, and litellm_logging_obj.complete_streaming_response,
all of which still pass through.
The conversation still reaches the scanner in full via `inputs`, so no
detection coverage changes.
Trimming happens before serialization, so the duplicate is never encoded.
Existing payload tests asserted the duplication; they now assert the trim
while keeping what they originally guarded - deep-copy semantics and the
unpicklable-object (uvloop.Loop) regression.
* 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.
* fix(panw_prisma_airs): scan tool call args as plain text, not a tool_event
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(panw_prisma_airs): type the tool call argument extractor
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(panw_prisma_airs): cover tool call error fallback and dict masking paths
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(panw_prisma_airs): scan tool names with args and tolerate custom tool calls
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(panw_prisma_airs): scan tool call arguments that arrive already parsed
The tool call slice types arguments as a string, so a client posting parsed JSON
failed validation and the whole tool call, name included, read as unscannable and
was skipped without ever reaching AIRS. The OpenAI request path forwards
client-supplied tool_calls verbatim, so that shape is reachable.
Coerce non-string arguments instead of rejecting them, so the content is scanned.
* fix(panw_prisma_airs): route tool-block masked data by scan side, not by key name
Merging #37036 (already on staging) with this PR produces no conflict and a
silent bug. #37036 withholds prompt_masked_data on response-side tool blocks,
which was right while tool calls went out as a request-side tool_event: AIRS
reported the model's arguments under that key. This PR scans tool calls as
ordinary prompt/response text, so the side of the scan now decides which key
holds what. The model's arguments arrive under response_masked_data, already
covered by _CLIENT_HIDDEN_SCAN_FIELDS, and prompt_masked_data goes back to
being the caller's own input -- one of the audit fields LIT-5638 asks for.
Left as merged, a response-side tool block drops that field with nothing to
flag it.
- Tool-path block branch calls _build_error_detail without also_hide
- also_hide parameter removed; after this change it has no callers
- Regression test asserts both directions: model output withheld, caller
input preserved. It fails against the auto-merged combination.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(panw_prisma_airs): a wrong-typed tool name must not suppress the scan
_ToolCallFunctionSlice types name as str, and _get_tool_call_function turns any
ValidationError into (None, None), which _scan_tool_calls_for_guardrail reads as
an unscannable tool call and skips. So a client posting "name": 123 keeps its
arguments off the wire to AIRS entirely -- no error, no log, no block. The
OpenAI request path forwards client tool_calls verbatim, so this is reachable by
any caller holding a valid key.
_coerce_arguments already existed for exactly this failure mode on the sibling
field. Widening it to cover name closes the gap:
name='transfer_funds' AIRS called: 1x args scanned: True
name=123 (int) AIRS called: 0x args scanned: False <- before
name=123 (int) AIRS called: 1x args scanned: True <- after
Reported by Cursor Bugbot on fd9f6396e5.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Yucheng Zhu <yucheng@berri.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(guardrails): return the full PANW AIRS scan response on blocked requests
The blocked-request error detail was assembled from a hardcoded allowlist, so audit fields like prompt_detection_details, prompt_masked_data, source, transaction_id and session_id never reached the client even though AIRS returned them.
Resolves LIT-5638
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(guardrails): drop redundant comment in AIRS error detail
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(panw_prisma_airs): withhold response_masked_data from the blocked-response error
The full AIRS passthrough also reached the response-side block path, where
response_masked_data carries the model's own generation. That branch is only
reached when mask_response_content is False, so the operator had explicitly
declined to deliver that text, and the error body handed it back anyway.
Withhold response_masked_data from the client-visible detail. prompt_masked_data
stays: it is the caller's own input and one of the fields the ticket asks for.
Every other AIRS field, including prompt_detection_details, source,
transaction_id and session_id, is unchanged.
* fix(panw_prisma_airs): withhold generated tool args from response-side blocks
_scan_tool_calls_for_guardrail calls AIRS with is_response=False because
tool_event is request-side in the AIRS schema, so AIRS returns the scanned
tool arguments under prompt_masked_data. When the tool calls being scanned
are the model's own output, that key holds generated content, and the
_CLIENT_HIDDEN_SCAN_FIELDS default (response_masked_data, empty on this
path) does not cover it. With the default mask_response_content=False the
block branch then shipped the model's masked tool arguments in the 400 --
the same content channel this PR closed for response_masked_data.
_build_error_detail takes an extra_hidden_fields argument so the withholding
stays in one place, and the tool-call block branch passes prompt_masked_data
when is_response is True. Request-side blocks are unchanged and still carry
prompt_masked_data, which is what LIT-5638 asks for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* style(panw_prisma_airs): apply ruff format
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>
Co-authored-by: Yucheng Zhu <yucheng@berri.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(panw_prisma_airs): surface scan_id and scan metadata on allowed requests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style: ruff format panw guardrail
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(panw_prisma_airs): expose scan id header only
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(panw_prisma_airs): inject http client instead of patching private api
Adds an http_client seam so the scan-id tests drive the real AIRS request/parse path through a mock transport, plus direct coverage for the scan-id header helper.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): expose guardrail scan id header to browser clients
Keeps the panw optional_fields block untouched to avoid a needless conflict with a sibling PR that deletes it.
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>
* fix(guardrails): scan and re-emit raw Anthropic SSE streams in the bedrock post-call hook
* fix(guardrails): keep upstream id and model on a blocked Anthropic stream
* fix(guardrails): deliver a blocked Anthropic stream as an error frame
* fix(guardrails): deliver an unscannable Anthropic stream as an error frame
* fix(guardrails): emit the guardrail block detail as JSON in the stream error frame
* fix(guardrails): deliver an Anthropic block through the shared block-SSE builder
* fix(guardrails): keep the shared SSE assembler behavior-identical for existing callers
* fix(guardrails): keep the stream error message a string and drop an unreachable branch
* chore(guardrails): drop a comment that repeated its own docstring
* fix(guardrails): let bedrock service failures keep their status instead of framing them as blocks
* fix(guardrails): key the streamed block decision on status, not detail shape
InvokeGuardrailChecks details a Mapping on its 500 for an unparseable response,
so a detail-shape test read that outage as a policy block and framed it as a 200
guardrail_error. Both block sites raise 400, so gate on the status too.
* refactor(guardrails): narrow the SSE error-frame helper to the input it actually takes
Both callers pass a string, so the Mapping overload and its json.dumps branch
were unreachable. Folds the block branch's narrative comment into the rebind
suppressions that already carry a reason.
Three groups, all verified by running the suite rather than by inspection.
18 files whose every test function carries an unconditional @pytest.mark.skip,
39 test functions in total. They are collected on every CI run and always skip,
so they advertise coverage the suite does not have. Reasons on the marks include
"AWS Suspended Account", "lakera deprecated their v1 endpoint" and "moved to
using 'otel' for logging"; 26 of the marks predate 2025.
30 test functions with a byte-identical body and identical decorators to a
sibling in the same file and class, differing only in name. Deleting one of each
pair removes no coverage. Four further candidates were excluded because they
override an inherited test, where deleting the override un-shadows the base
class implementation instead of removing a duplicate.
9 test functions that a later definition of the same name shadows, so Python
never binds them and pytest cannot collect them.
One file that is a demo script rather than a test; its own docstring says to run
it with python.
Verification: collecting the 26 edited files gives 2,492 node IDs before and
2,462 after. The 30 duplicate deletions account for exactly 30 removals, the 9
shadowed deletions account for 0 (confirming at runtime that they were never
collectable), nothing unexplained disappeared, and nothing new appeared. No
other test or module imports any deleted symbol.
* feat(guardrails): chunk oversized Bedrock ApplyGuardrail requests instead of failing
AWS's ApplyGuardrail API rejects requests whose content exceeds the
account's per-request "maximum input size in text units" quota with a
400 ValidationException. That cap is account/region/policy-dependent
and cannot be predicted from config, so it can only be reacted to.
_make_apply_guardrail_request now tries the whole-content call first
(no behavior change for requests that already fit). On a too-large
ValidationException it bisects the flat content list and retries each
half sequentially, recursing until every piece fits or cannot be split
further, then merges the per-chunk responses (action, assessments,
outputs, usage) into one so callers cannot tell chunking happened. A
real guardrail block on any (sub-)chunk still raises immediately.
Contextual-grounding requests are never chunked: grounding scores the
response holistically against the whole reference source, so
fragmenting it would produce misleading scores.
Each chunk call also gets a small exponential backoff retry on AWS
ThrottlingException (429), since chunking increases the number of
per-second API calls and can trade a 400 for a 429.
All new state is local to a single request's call stack (no shared
cache, no cross-process coordination), so this is safe for
single-pod, multi-pod, and cache-less LiteLLM proxy deployments alike.
* fix(guardrails): address Bedrock ApplyGuardrail chunking review feedback
Fixes three issues flagged in review of the chunking fallback: a single
oversized content item couldn't be split (only list-length bisection was
supported), a chunked request that got recovered still logged a stray
failure telemetry entry alongside the real outcome, and flattening chunk
outputs without positional bookkeeping could misalign masked text onto
the wrong original message once a chunk had nothing to mask.
* test(guardrails): add regression test for multi-level Bedrock guardrail chunking
Confirms the too-large bisection recursion isn't capped at a single split:
a payload that is still oversized after the first halving keeps splitting
until every piece fits, converging on however many chunks it takes rather
than only ever producing two.
* fix(guardrails): hybrid bin-pack+bisection chunking, whitespace-safe splits
Rework Bedrock ApplyGuardrail chunking from pure reactive bisection to a
hybrid strategy: bin-pack content into fixed-budget batches up front as
the fast path, falling back to the existing recursive bisection only for
a batch AWS still rejects as too large. Avoids paying O(log n) round
trips on every oversized request when a single pass would do.
Also switch single-item text splitting from a raw character midpoint to
the nearest whitespace boundary, so a fragment never starts or ends
mid-word. Closes the accidental-severing case from review; the residual
gap (a multi-word denied phrase deliberately straddling the boundary) is
documented as an accepted limitation, since fixing it would require an
overlap window reconciled against masked output with no documented
length-preservation guarantee from AWS.
* chore(ui): regenerate dashboard API types
* fix(guardrails): don't retry an oversized Bedrock guardrail call as a throttle
AWS reports an ApplyGuardrail request that exceeds the per-request
text-unit cap as a ThrottlingException (429), not only as the documented
ValidationException (400). Verified against a live guardrail with an
active content-filter policy: a 3273-text-unit request comes back as
"Input text size (3273 text units) exceeds the maximum allowed (1000 text
units) for the content filter policy (Classic tier)".
The throttle retry keyed off status 429 alone, so every oversized chunk
burned the full backoff-retry budget - each attempt a billed AWS call
preceded by a sleep - before the bisection fallback got a chance, at every
level of the recursion. A size error is not transient; re-posting the same
content can never succeed. It now short-circuits straight to bisection.
Also rename _is_input_too_large_validation_error to
_is_input_too_large_error (it never keyed off the status code, and the
error is not always a ValidationException), correct the docstrings that
asserted a 400, and log at warning level when a split happens so the
recovery is visible without --detailed_debug.
* Revert "chore(ui): regenerate dashboard API types"
This reverts commit ebf8ba2fd5.
* fix(guardrails): group all fragments of one item and stop double-logging
Two defects found in review, both invisible to the existing tests.
Fragment grouping assumed a split content item always produces exactly two
adjacent fragments. That holds for one bisection level but not two: an item
split twice yields four fragments, which were regrouped in fixed pairs into
two output entries for a single message. Since masking walks the merged
outputs by a running index across the original, unchunked message list, that
message was written back truncated to its first half and every later message
shifted. Fragments now carry the size of the group they belong to, so any
number of them collapse back into exactly one output entry.
Telemetry was also double-counted. AsyncHTTPHandler.post calls
raise_for_status(), so every non-200 from Bedrock reaches _sign_and_post's
error path, which logged guardrail_failed_to_respond before re-raising as an
HTTPException that the consolidating caller then logged again. A request
recovered by chunking reported one failure per rejected attempt plus a
success. The ApplyGuardrail path now opts out of that per-attempt logging,
since it owns consolidated per-request logging; the connection-level branch
still logs, as nothing else records it.
The existing tests missed both because their mocks return a non-200 response
object, while the real client raises. Added a helper that raises a genuine
httpx.HTTPStatusError so these paths are covered the way production hits
them, plus a case asserting an unrecoverable failure still logs exactly once
rather than zero times.
* refactor(guardrails): move Bedrock chunking rationale into docstrings
The chunking work explained itself with inline comment blocks, which this
repo's conventions do not want. Folded that reasoning into the docstrings of
the functions it describes and dropped the comments, including the
module-level constant blocks and the test-file banner.
No behavior change. The banner also claimed AWS rejects an oversized request
with a 400 ValidationException, which live testing disproved, so removing it
drops a stale claim as well as an internal ticket reference from a public repo.
* feat(guardrails): match AWS default chunk budget and make it configurable
ApplyGuardrail's default quota is 25 text units, roughly 25,000 characters,
per second. Chunking has to respect that throughput limit rather than just the
per-request size, otherwise splitting an oversized request trades a size error
for a throttle. The budget now defaults to 25,000 to match that default for
every user, up from an arbitrary 20,000.
Accounts with raised quotas can spend fewer calls by setting
chunk_budget_chars on the guardrail. A value AWS still rejects as too large is
bisected automatically, so an over-large setting costs an extra round trip
rather than failing the request.
* fix(guardrails): never split a Bedrock text into an empty fragment
_nearest_whitespace_split_index could return len(text) when the only space at
or after the midpoint was the final character, so the first fragment came back
identical to the text AWS had just rejected as too large and the second came
back empty. AWS rejects the unchanged fragment again, and each retry re-splits
it into the same fragment, so an oversized single item shaped like a long
unbroken token with one trailing space exhausted the stack with a
RecursionError instead of scanning or surfacing Bedrock's error.
Candidate boundaries that would leave either side empty are now discarded, and
the raw midpoint is used when none remain. The midpoint is always safe because
_split_bedrock_content only calls this for text of at least two characters.
* style(guardrails): move chunking rationale out of comments and into docstrings
* fix(guardrails): raise 500 when Bedrock reports a failure inside a 200 body
Also types the credentials parameter on the new chunking helpers and rebuilds
fragment grouping without mutating a list or rebinding an index
* fix(guardrails): raise 500 when Bedrock reports a failure inside a 200 body
Restores the source changes intended for a08e4cf309, which landed with only the
test. Also types the credentials parameter on the new chunking helpers and rebuilds
fragment grouping without mutating a list or rebinding an index
* style(guardrails): sort the constants import into the first-party block
* refactor(guardrails): bring the Bedrock chunking path under the LIT lint budgets
Annotates never-rebound locals with Final, replaces the retry counter and the two
branch-assigned locals with single bindings, and moves the internal chunking chain
to Sequence parameters and tuple returns. Collections that reach the logged payload
stay lists on purpose: redact_nested_match_and_regex_keys only traverses dict and
list, so a tuple would carry PII past redaction. The remaining constructions are
contract-bound and carry inline reasons
* fix(guardrails): keep the pre-chunking contract for failures reported inside a 200
Reverts the 500 this branch introduced for an AWS 200 whose body carries an
Output.__type exception marker: the request proceeds as it did before chunking
existed. The logged status is now derived from the merged response instead of
being hardcoded to success, so that shape is still reported as
guardrail_failed_to_respond. The consolidated failure logger also goes back to
logging a dict rather than a bare string, matching both the pre-chunking code and
the InvokeGuardrailChecks path in this file
* docs(guardrails): correct the docstring for failures reported inside a 200 body
The raise was reverted, so the docstring no longer describes the code. Records that
the request proceeds by design and points at LIT-5338 for closing the fail-open path
behind the existing unreachable_fallback setting
---------
Co-authored-by: spencer-burridge <265588760+spencer-burridge@users.noreply.github.com>
Prompt Security drops tool and function rows unless check_tool_results
is on, so it now reports scan-only support from that setting and the
registry refuses the pairing at boot. Pairing scan_only_tool_results
with skip_tool_message_in_guardrail excludes every message, so guardrail
initialization now rejects that combination too.
An SSE stream that cannot be positively identified as Anthropic (no
parseable message_start event) now blocks instead of passing through
unscanned, closing the bypass where any raw-SSE backend skipped tool
permission checks entirely. Buffered chunks are joined back into one
stream before parsing, so events split across network chunk boundaries
assemble correctly instead of being silently dropped. Rewrite mode now
resets finish_reason to stop when no tool call survives, so the
re-encoded Anthropic stream reports stop_reason end_turn and clients do
not wait for a tool result that never comes
Guardrails silently skipped three surfaces on the Anthropic Messages
path, so an agent loop driven by /v1/messages ran unguarded:
- The Anthropic input translation never walked tool_result blocks, so
content returned by a local tool (a curl, a file read, an MCP call)
reached the model unscanned in both the string and list content
shapes, images inside a tool_result included.
- tool_permission only understood ModelResponse, so an Anthropic
non-streaming response or a raw SSE stream carrying tool_use blocks
passed through with no rule ever evaluated.
- ContentFilterGuardrail scanned inputs["texts"] but never
inputs["tool_calls"], so the arguments a model proposes for a tool
call went unchecked.
Tool call arguments are parsed as JSON before filtering so a MASK
action rewrites the value and leaves the payload valid JSON; non-JSON
arguments fall back to scanning the raw string. Denied tool_use blocks
are dropped from the Anthropic content array and replaced with a text
block, and stop_reason resets to end_turn when nothing tool-shaped
survives.
ContentFilterGuardrail implements apply_guardrail, which is everything the
generic post_mcp_call_hook machinery needs to scan an MCP tool result before
it reaches the model, but post_mcp_call was missing from
get_supported_event_hooks. _validate_event_hook rejects any mode outside that
list, so a config with `mode: post_mcp_call` failed proxy startup with
"Event hook GuardrailEventHooks.post_mcp_call is not in the supported event
hooks" instead of scanning tool output.
Declaring the hook makes the indirect-prompt-injection case enforceable: an
MCP fetch tool returns a page whose body carries "IGNORE ALL PREVIOUS
INSTRUCTIONS ...", and the gateway blocks the result rather than handing it
to the model.
* fix(proxy): give proxy_admin_viewer read parity with proxy_admin
Route-level checks already default-allow management GETs for the viewer
role, but ~15 handlers compared user_role to PROXY_ADMIN only, dropping
viewers into regular-user scoping (/key/list, /user/info, /model/info,
guardrails, prompts, agents, memory, workflows, MCP catalog, coordination
redis settings, credential migration check, enterprise projects). Swap
those read paths to user_api_key_has_admin_view; write gates unchanged.
The dashboard now presents the viewer session as Admin for all gating
(effectiveSessionRole) so every page fetches with admin visibility, with
userRoleLabel/isViewOnly preserving the account-menu label and the
playground cost guard. The server remains the write authority.
* refactor(agents): remove side-effectful health_check param from GET /v1/agents
Addresses a security review finding on the admin viewer read parity change:
listing agents with health_check=true made the proxy issue a server-side GET
to every agent URL, so a read-scoped caller could trigger request fan-out
beyond their object permissions. The list endpoint is now a pure read for
every role.
Removes the query param, the URL probing helper and its timeouts, the
AgentHealthCheck httpx provider tag, and the dashboard's Health Check
toggle. Requests still passing health_check=true get the full list back
with the param ignored.
* fix(proxy): keep credential encryption check proxy_admin only
The residual scan behind GET /credentials/migrate-encryption/check loads
every model, credential, MCP, team, and verification-token row and runs a
decryption attempt on each stored value. Extending it to proxy_admin_viewer
let a read-only account repeatedly trigger deployment-wide scans, so the
route keeps its original full-admin gate.
* fix(agents): restore health_check, keep list fast path proxy_admin only
Restores the agent health_check feature exactly as before this PR: the
query param, the URL probing helper, the httpx provider tag, and the
dashboard toggle all return, so existing callers keep the filtering
contract. The viewer expansion is instead reverted at its source: the
GET /v1/agents admin fast path stays PROXY_ADMIN only, so a
proxy_admin_viewer goes through the object-permission scoped branch as
before and cannot fan out health checks beyond their allowlist. The
viewer read of a single agent stays viewer-inclusive since it has no
side effects.
ruff.toml has excluded litellm/types/* since 2024, so no lint rule ever ran
on the types tree. Remove the exclusion, apply ruff --fix and ruff format
across litellm/types, and hand-fix what autofix cannot reach so the
pyupgrade budgets stay at zero: implicit type aliases converted to PEP 604
unions, RootModel[Union[...]] bases, duplicate imports, and a stray print.
Load-bearing import X as X re-exports deleted by preview-mode F401 are
restored, and the six star-imported hub modules keep their re-export
surface via per-file F401 ignores. Star-import consumers that silently
relied on typing names leaking from those hubs are modernized to builtin
generics and PEP 604 unions.
Runtime annotation introspection that only recognized typing.Union is
taught types.UnionType (guardrail UI field schemas, volcengine response
fill), with regression tests for both. Strict budget limits for the rules
the types tree now trips are raised to exact measured totals, so any
net-new violation still fails the gate
The proxy serves POST /openai/v1/responses alongside /responses and
/v1/responses, but only the latter two were in API_ROUTE_TO_CALL_TYPES.
UnifiedLLMGuardrails.async_post_call_success_hook resolves the call type
from request_route, so on the alias it resolved to None and returned the
response unscanned; model output reached the client with post-call
guardrails never running. The key and team tool allowlist was unenforced
on the same alias for the same reason.
Register the alias family in API_ROUTE_TO_CALL_TYPES and in
LiteLLMRoutes.openai_routes, mirroring how the /openai/v1/realtime
aliases are registered, and log a warning at the two points where the
unified guardrail skips post-call scanning so a future unmapped route is
visible instead of silent.
The Responses block of API_ROUTE_TO_CALL_TYPES moves from list to tuple
literals because the LIT002 budget rejects net-new mutable-collection
construction; the map is read-only, so it is now typed as a Mapping of
Sequence and the budgets ratchet down accordingly.
A bedrock guardrail configured mode: during_mcp_call never ran. ProxyLogging
remapped the event to during_mcp_call and dispatched, but bedrock's own
async_moderation_hook then hard-coded during_call and re-checked, so the second
check rejected the very requests the guardrail was configured for and the tool
call proceeded unscanned with no error.
Remap call_mcp_tool the way model_armor already does, which matches the remap
ProxyLogging.during_call_hook itself performs, and teach the shared
get_guardrails_messages_for_call_type helper that an MCP tool call carries its
payload in the same messages key, without which the hook passes the gate and
then bails on an empty message list.
The Headroom guardrail sent every message to /v1/compress, including the
system prompt and the user's current instruction. On an agentic /v1/messages
request the live turn is the largest compressible blob, so it came back as a
hash marker; the model then called headroom_retrieve and got its own
instruction returned in a tool_result block, which reads as data it fetched
rather than a request to act on, so it described the content instead of doing
the work.
litellm already owns the policy for what a compressor may never rewrite:
get_protected_indices covers the system rows, the last user row and the last
assistant row, and compress() expands it over whole tool exchanges. Headroom
now consults it (promoted from a private name and given tests) and expands it
the same way, so the trailing tool result cannot come back as a marker
standing in for the result of the call the model just made. Protected rows are
withheld from the payload rather than pinned afterwards, so their tokens are
not reported as savings that are never applied; the write-back discards a
compressed system prompt outright, so that saving never existed. The cost is
that a query-aware service no longer sees the newest user message.
A response whose row count differs from what was sent can no longer be
interleaved with the withheld rows, so it goes through the configured fail
policy instead of being adopted. Fail-open now returns the caller's own inputs
object: translation handlers detect a rewrite by identity, so a rebuilt copy
sent an unchanged request through the Anthropic write-back for nothing.
That write-back rebuilt the request with one anthropic_messages_pt call, which
merges every run of consecutive user/tool rows, so a tool_result turn and the
user turn after it arrived fused. Converting a row at a time would separate
them but breaks tool pairing: with modify_params on, an assistant row whose
results are converted separately reads as an orphaned tool call and the
sanitizer answers it with a synthetic "tool execution skipped" result while
dropping the real one. Conversion is now grouped by tool_call_id ownership,
which satisfies both, and the same grouping decides which rows headroom
protects, so the two agree by construction.
The CCR follow-up also dropped any text the model wrote alongside its tool
call, and echoed tool calls it had no results for. Both are fixed by reusing
compresr's extraction helper, now shared instead of duplicated.
Resolves LIT-5018
Anthropic-format requests translate to messages whose content is a list
of part dicts, which the headroom compression service's transforms
silently skip (they only rewrite string content), so compression never
applied to Anthropic client traffic while the guardrail still reported
itself as applied.
Flatten all-text part lists to plain strings for /v1/compress and
restore the original shapes from the response: untouched rows keep
their exact original parts, a rewritten row collapses to one part
carrying the last declared cache_control breakpoint (a breakpoint
caches the prefix ending at its part, so the last one and its TTL
still describe the merged row). Rows with any non-text part are never
flattened, since merging text across a non-text part would move a
later breakpoint to the other side of it; they pass through the
service untouched, matching its own behavior for non-string content.
Flattening and write-back use the shared content_text helpers that
compresr's breakpoint fix also uses.
Resolves LIT-4795
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(guardrails): resolve judge_model credentials via Router in llm_as_a_judge
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): wire llm_router into DB-backed judge guardrail init paths
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(guardrails): assert patch endpoint forwards llm_router to sync
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(guardrails): resolve judge Router lazily and fix wildcard/alias dispatch
Resolve the proxy Router at judge-call time via an injected provider instead of
capturing it at construction, so a DB-backed judge guardrail created before the
Router exists no longer captures None permanently. Select the Router path with
router.get_model_list(model_name=judge_model) so wildcard routes and
model_group_alias keys resolve, not just literal deployment names. Isolate the
judge call from user-traffic routing with num_retries=0 and fallbacks=[].
Revert the llm_router threading through the DB sync/reinit/create/approve/patch
paths since the lazy provider makes it unnecessary. Replace mocked-Router tests
with real Router coverage for plain deployments, model_group_alias, and wildcard
routes, plus lazy per-call resolution.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): harden judge verdict parsing and guard proxy import
Strip markdown fences and surrounding prose before json.loads so fencing-prone
judge models evaluate instead of failing open, guard the proxy_server import in
_default_router_provider so an unimportable proxy falls back to the SDK, and
snapshot/restore global callback lists in the DB-path judge registry tests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): reject non-object judge verdicts instead of failing open as success
* fix(guardrails): route hidden model_group_alias judge models through the Router
---------
Co-authored-by: milan <milan@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: yucheng-berri <yucheng@berri.ai>
Anthropic cache_control breakpoints are positional: each one caches the
prefix ending at the part that carries it. Compresr flattened every text
part of a message into one string and wrote the compressed result back
into the first text part only, which dropped every later breakpoint and,
when a non-text part sat between text parts, moved the trailing text to
the other side of it.
The positional invariant now has one owner. guardrail_hooks/content_text.py
holds content_to_text alongside is_all_text_parts and
merge_rewritten_text_parts, so a compressed string is only ever written
back over a contiguous run of text parts, and the merged part carries the
last declared breakpoint and its TTL.
Compresr consumes that owner at both ends: _select_targets no longer
selects a row holding a non-text part, and _replace_text_in_content
returns such a row unchanged rather than merging across it. Rows whose
content is a plain string are unaffected.
Mixed rows therefore stop being compressed, which is a deliberate trade;
no single-string write-back can preserve a breakpoint across a non-text
part, so the alternative is silently caching a different prefix than the
caller configured.
* fix(guardrails): add /v1/messages support for Straiker plugin
- Pass prepared response data to Anthropic Messages streaming post-call hooks (litellm/llms/anthropic/chat/guardrail_translation/handler.py)
- Normalize Straiker request, tool, finish-reason, and mode fields across Chat Completions, Messages, and Responses APIs
* fix(guardrails): gate cross-surface message resolution and cover streaming request data
Resolve request messages only for surfaces that have a mapped translation
handler. The unguarded fallback tried every registered handler in turn, which
raised AttributeError out of the guardrail's error handling on list-shaped
`input` bodies, and synthesized a chat message that was never sent for bodies
it happened to parse.
Prepare request data on the mid-stream Anthropic branch as well, matching the
terminal branch and the OpenAI handler, so guardrails that scan before
end-of-stream still receive identity metadata.
Read usage from Anthropic dict responses so non-streaming /v1/messages reports
token counts instead of null.
Add regression coverage for the streaming request data on both the terminal and
mid-stream branches; reverting either now fails.
---------
Co-authored-by: cs-mehta <chandra@straiker.ai>
The savings readers (extract_compression_saved_tokens, feeding
compression_saved_tokens on the daily spend tables) key exclusively on
tokens_saved in the guardrail_response stats, but the Headroom guardrail
builds those stats as a filtered pass-through of the compression service
response and the live service omits tokens_saved. Every compressed request
recorded 0 saved tokens on the Cost Optimization dashboard.
Derive tokens_saved = tokens_before - tokens_after when the key is absent
and both operands are numeric; a service-sent value still wins. The two
sibling writers (compresr, native compression interception) already derive
it the same way.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
