Runtime regression in the previous simplification — the guardrail
registry iterates every registered class at load time and calls
get_supported_event_hooks(). Fallback of ConductGuardrail = None
crashed the whole registry with AttributeError, which cascaded into
unrelated guardrails' tests (noma_v2, repelloai, hide_secrets,
provider_specific_params, etc.).
Fallback now defines ConductGuardrail as a real subclass of
CustomGuardrail with the required class attrs (SUPPORTED_EVENT_HOOKS
+ get_supported_event_hooks). Matches the pattern the
guardrails_ai integration already uses in the same repo.
raise_if_missing_package still fires before instantiation so users
see the friendly pip install error.
All four budget gates re-verified locally against
upstream/litellm_internal_staging:
ruff_strict_gate OK
test_quality_gate OK
type_discipline_gate OK
type_check_gate OK
Plugin conduct-litellm-guard 0.2.3 ships SUPPORTED_EVENT_HOOKS +
get_supported_event_hooks on ConductGuard directly. The upstream
shim's subclass wrapper is now redundant — dropping it clears every
strict-rule budget gate (ruff-strict / test-quality /
type-discipline / basedpyright) in one pass.
Changes:
- conduct.py: subclass removed; ConductGuardrail is now an alias for
the plugin's ConductGuard (no dynamic base class, no reassignment,
no # type: ignore). raise_if_missing_package helper unchanged.
- test file: _IMPORT_ERROR → _import_error rename to satisfy
reportConstantRedefinition (basedpyright treats SCREAMING_CASE as
constant). Also drops unused sys import.
- Pin bumped to conduct-litellm-guard>=0.2.3 in the module docstring.
Verified all four LiteLLM gate scripts locally against
upstream/litellm_internal_staging:
ruff_strict_gate OK
test_quality_gate OK
type_discipline_gate OK
type_check_gate OK
BerriAI/litellm CI's test-quality budget flagged
test_raise_if_missing_package_is_noop_when_present as a zero-assert
test (TQ001). Make the intent explicit: raise_if_missing_package()
must return None when the package IS installed.
BerriAI/litellm CI's type-discipline budget check flagged the
subclass __init__ shim. Fixes:
- Drop the __init__ override entirely — the subclass now inherits
__init__ from _BaseConductGuard (when the standalone package is
installed) or from CustomGuardrail (fallback). Removes both the
banned **kwargs (LIT008) and all four inert # type: ignore markers
(LIT009 x 4).
- Move the missing-package check into a dedicated
raise_if_missing_package() helper called by
initialize_guardrail before construction. Preserves the
cursor[bot] fix (silent-drop-on-import-failure) without needing
a custom __init__.
- Fallback branch aliases _BaseConductGuard = CustomGuardrail
directly, no type-ignore comment needed.
- Test updated to exercise the helper instead of the removed
__init__ path; new companion test asserts the helper is a no-op
when the package IS installed.
Local: ruff --select ANN,TID passes clean. ruff format applied.
Same behavioral surface — user-visible error message unchanged.
BerriAI/litellm CI's ruff strict-rule budget check flagged four new
violations on the conduct shim. Fixes:
- __init__.py: add return type annotation on initialize_guardrail
(ANN201)
- conduct.py: swap Any → object on __init__(*args, **kwargs) so the
signature stays permissive without dynamically-typed Any (ANN401)
- conduct.py: drop the now-unused Any import (TID251)
Ruff --select ANN,TID passes locally.
Rename fail_mode → unreachable_fallback (typed field)
─────────────────────────────────────────────────────
The shim was reading a free-form ``fail_mode`` field; a typo silently
defaulted the plugin to fail-open behavior. Switch to the typed
``LitellmParams.unreachable_fallback`` field so Pydantic validates the
value at config load. The plugin's constructor kwarg stays as
``fail_mode`` — the initializer maps the typed field onto it.
(yucheng-berri, devin-ai-integration)
Fix timeout default (was silently discarded)
────────────────────────────────────────────
``getattr(litellm_params, "timeout", 8.0)`` only applied the default
when the attribute was missing; ``LitellmParams.timeout`` always
exists and defaults to ``None``, so the intended 8-second budget was
never used. Change to ``getattr(..., None) or 8.0`` so ``None`` (and
``0``) fall through to the default.
(cursor[bot])
Move ImportError from module-load to __init__
─────────────────────────────────────────────
Raising ImportError at module load caused the guardrail-hook
auto-discovery loop to silently drop the registration when
``conduct-litellm-guard`` was missing. Users saw configs load with
no guardrail active and no error. Import lazily; raise the friendly
``pip install`` error at ``ConductGuardrail.__init__`` when
actionable.
(cursor[bot])
Advertise only supported event hooks
────────────────────────────────────
``during_call`` mode was advertised in the guardrail config but the
class never overrode ``async_moderation_hook`` — every request in that
mode silently bypassed policy. Override ``get_supported_event_hooks``
to return only ``pre_call`` so LiteLLM validates configs against
supported modes at load time. ``during_call`` / ``post_call`` support
lands with plugin 0.3.x once the underlying response-gate is wired
through ``guard_check_response``.
(veria-ai)
Text-completion + full-turn prompt scanning
───────────────────────────────────────────
Fixed in the standalone package: ``conduct-litellm-guard 0.2.2``
(BerriAI/litellm PR #38143 companion, shipping to PyPI shortly).
Pinned in the docstring here as the minimum supported version.
(veria-ai — text_completion bypass + 4KB truncation)
Tests
─────
* ``test_only_pre_call_event_hook_advertised`` — regression for
``during_call`` silent-bypass finding
* ``test_initialize_prefers_typed_unreachable_fallback`` — regression
for typo silent-fail-open finding
* ``test_initialize_applies_timeout_default_when_field_is_none`` —
regression for silently-discarded 8.0 default
* ``test_missing_standalone_package_raises_at_construction`` —
regression for silent-drop-on-import-failure finding (previous
module-load raise replaced with lazy import + init-time raise)
Codecov flagged the __init__.initialize_guardrail body as uncovered
(30% patch coverage on that file). Added a test that mocks
litellm.logging_callback_manager and calls initialize_guardrail with
a SimpleNamespace stand-in for LitellmParams — exercises the full
function body and confirms the callback is registered.
The wrapper module imports its runtime from the conduct-litellm-guard
PyPI package. When the package is not installed in the CI environment,
the smoke tests can't verify wiring (the import raises before any test
runs). Use pytest.importorskip so BerriAI's default CI env doesn't
fail on this integration, while environments that do install the
package (via 'pip install conduct-litellm-guard[dev]' or similar)
still get the smoke coverage.
Full behavioural test coverage lives in the conduct-litellm-guard
package's own CI.
The full adapter (response parser, session-ID chain, fail-mode logic,
HTTP client) lives in the conduct-litellm-guard package on PyPI. The
upstream tree hosts a thin re-export + the LiteLLM registration wiring.
Matches the Aporia / Lakera pattern — vendor SDK on PyPI, upstream
integration is a tiny adapter.
Benefits:
- Passes ruff-strict-budget and type-discipline-budget without new
violations.
- Users get the same install experience as any other guardrail vendor:
pip install conduct-litellm-guard
- Vendor keeps ownership of the parser + fail-mode semantics; upstream
keeps a stable interface.
Tests slimmed to smoke coverage (imports work, class is a
CustomGuardrail, enum + registries wired, missing-package error path).
Full behavioural coverage stays in the PyPI package.
Local runs of both scripts/ruff_strict_gate.py and
scripts/type_discipline_gate.py against upstream/litellm_internal_staging:
both pass.
Removing 'Any' in the prior commit left **kwargs untyped, which
tripped ANN003 (missing type annotation on **kwargs). Using 'object'
threads the strict-rule budget cleanly.
BerriAI's ruff strict-rule budget caps ANN401 (Any type annotation)
and TID251 (banned import) totals. Aligning with the CustomLogger
base signature (data: dict, cache: object, **kwargs untyped)
eliminates all Any uses in the module. Local tests still pass 15/15.
Adds Conduct Guard as a first-class LiteLLM guardrail. Point any
LiteLLM proxy at Conduct and every LLM call routed through it is
policy-checked before the upstream request goes out — block, warn,
audit, or trigger a human-in-the-loop approval, with the same signed
configuration + hash-chained audit log Conduct exposes on its native
enforcement surfaces.
- litellm/types/guardrails.py: add CONDUCT to SupportedGuardrailIntegrations.
- litellm/proxy/guardrails/guardrail_hooks/conduct/__init__.py: registration
via guardrail_initializer_registry and guardrail_class_registry, picked
up by the auto-discovery in guardrail_registry.py.
- litellm/proxy/guardrails/guardrail_hooks/conduct/conduct.py: the adapter.
CustomGuardrail subclass, async_pre_call_hook, response envelope parser
for the five Conduct verdicts (ok / advisory / WARNING / BLOCKED /
PENDING approval), fail-mode logic, session-ID resolution chain
(litellm_metadata.trace_id → X-Conduct-Session-Id → hash fallback).
- tests/test_litellm/proxy/guardrails/test_conduct_guardrail.py: envelope
parsing, pre-call allow/block/approval, config precedence, missing-token
construction error.
```yaml
guardrails:
- guardrail_name: conduct-guard
litellm_params:
guardrail: conduct
mode: pre_call
api_base: https://api.conductai.ai # optional, default
api_key: os.environ/CONDUCT_AGENT_TOKEN # cond_agt_* token
fail_mode: fail_closed # or fail_open
tool_name: llm_call # scoped tool_name
```
A standalone PyPI package `conduct-litellm-guard` shipped ahead of this
PR for teams pinned to older LiteLLM versions. Once this integration
merges, the standalone README will point at the native support as the
preferred path.
- PyPI: https://pypi.org/project/conduct-litellm-guard/
- Product: https://conductai.ai/guard
Contact: sudhi@b2bsphere.com
* fix(datadog_llm_obs): keep tool call and result structure under redaction and emit tool output tokens
Under datadog_llm_observability_params.turn_off_message_logging the span kept only one role plus "redacted-by-litellm" per message, so Datadog showed Tool Call 0, Tool Result 0 and no tool output token data. The shared CustomLogger hook collapsed the messages before the callback ran, and the Datadog redaction then dropped tool_calls and tool_results.
The Datadog callback now opts out of the shared message collapse (redacts_messages_itself) and redacts its own normalized messages, keeping roles, tool names, ids and types while replacing content, arguments and results. Tool result tokens are counted with litellm.token_counter before redaction and shipped as the tool_output_tokens metric. Other callbacks keep the inherited behavior.
Resolves LIT-7545
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(datadog_llm_obs): drop explanatory docstrings from the redaction change
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(ui): regenerate schema.d.ts for the classifier descriptions changed in #40655
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): write failure spend log for guardrail-blocked /mcp-rest/tools/call
call_tool_rest_api only translated exceptions to HTTP responses, so a pre_mcp_call
guardrail block never reached failure_handler / async_failure_handler /
post_call_failure_hook and no LiteLLM_SpendLogs failure row was written. Extract
the failure logging from call_mcp_tool into _fire_mcp_tool_call_failure_logging
and run it in the REST route for anything raised between
common_processing_pre_call_logic and execute_mcp_tool
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): keep the original REST tool error when failure logging raises
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): log virtual mcp_tool_call failures and keep REST success latency scoped to tool execution
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): keep call_type and request start time on failed-request spend logs
post_call_failure_hook pops litellm_logging_obj before the failure callbacks
run, so the spend row built from request_data had a blank call_type and used
datetime.now() as the start time. A guardrail-blocked MCP tool call therefore
showed up in the Logs page as an LLM row with no call type and a 0s duration.
Lift call_type and start_time off the logging object alongside the fields
already lifted, and have the DB failure hook prefer the lifted start time.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(proxy): inject the spend writer into _ProxyDBLogger instead of patching a module global
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yucheng <yucheng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(router): fall back from unhealthy auto-router tier
Co-Authored-By: Claude Code <noreply@anthropic.com>
(cherry picked from commit 00c7fd8376)
* fix(router): treat budget and tag exhaustion as a no-capacity verdict
The eligibility probe only read typed router errors as "nothing here can
serve this". Provider and deployment budget exhaustion, and tag routing
with no matching deployment, report it as a bare ValueError carrying a
RouterErrors marker, so the probe read a spent tier as live, skipped the
peer and default recovery, and failed the request.
---------
Co-authored-by: Tin Chi Lo <tin@berri.ai>
Co-authored-by: Claude Code <noreply@anthropic.com>
* fix(proxy): authorize every Responses API id, not only the ones the proxy issued
The ownership check on the Responses API only ran when the id arrived in the
proxy's own encrypted format. An id in any other shape skipped the check and
was forwarded upstream, so a key that did not own the response could retrieve,
cancel, delete, or chain off it.
Every addressed id now goes through one authorization step shared by retrieve,
cancel, delete, list-input-items, and create's previous_response_id. An id the
proxy did not issue is refused with 403 unless the deployment opts in with
general_settings.allow_unmanaged_response_ids, has responses id security
disabled, has no signing key configured, or the caller is a proxy admin.
* fix(proxy): re-authorize the retained responses id instead of trusting it
The docstring cited session_affinity_ttl_seconds as the keepalive bound,
but the Router-level knob feeding ttl_seconds is
deployment_affinity_ttl_seconds; session_affinity_ttl_seconds is the
separate per-request PreRoutingHookResponse override. Anyone grepping
the docstring's name to shrink the Router default finds only the
override. Name both, scoped correctly.
Starlette scans the route table in registration order, so a request pays one
regex match per route registered ahead of its own. The proxy registers several
hundred routes and left the liveness probe near position 280 and the lazy
loaded /v1/messages at the very end. Move /health/liveliness, /health/liveness,
/v1/chat/completions, /chat/completions and /v1/messages to the front of the
route table after startup registration and again after a lazy router loads.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(proxy): make the in-memory management cache capacity configurable
Add general_settings.user_api_key_cache_max_size (positive int, default 200) to resize the
in-memory tier of the shared user_api_key_cache at startup and on DB config reloads, expose it
in the Admin UI general settings, and cover it with behavioral tests. Prior art: #34726
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(caching): resize the in-memory tier from DualCache so any cache instance honours the cap
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(proxy): wrap the cache capacity field description to the 120 col limit
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>
* fix(passthrough): parse Bedrock stream spend incrementally instead of buffering the whole response
Bedrock pass-through streaming kept every relayed chunk in memory until EOF and
then decoded, parsed and translated the whole stream again for spend logging.
Large or concurrent streams could exhaust proxy worker memory.
Sync and async passthrough wrappers now hand each chunk to a provider stream
collector as it is relayed. Bedrock decodes event-stream frames incrementally,
folds consecutive text deltas, and keeps only what stream_chunk_builder needs
for usage, tool calls and metadata. Text deltas are no longer retained in the
Bedrock and Anthropic stream decoders either. Providers without a collector
keep the previous raw-bytes behavior. Collector failures are isolated so spend
tracking can never interrupt the customer stream
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(passthrough): assert the spend payload the collector builds instead of mock internals
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(passthrough): type the Bedrock collector helpers by the collector protocol instead of asserting the class
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>
Key objects share the 200-entry UserApiKeyCache in-memory store with teams,
end users, tags and memberships, so churn in those objects evicts hot keys
and forces a LiteLLM_VerificationToken lookup on the next request. Route
bare hashed-token keys to a dedicated InMemoryCache inside UserApiKeyCache
while keeping Redis, TTL, serialization and invalidation shared
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
