The post-call hook was hardcoding tool_results=[] on every Turn, so the
failure detector never saw tool errors and the bandit only learned from
satisfaction — never from negative tool outcomes.
Added _recent_tool_results(messages): walks the request messages from the
tail and collects the contiguous run of role=='tool' entries — those are
the results from the most recent assistant tool_calls round. Normalizes
each to {content, is_error}, the only fields signals._detect_failure /
_detect_exhaustion read.
Tests: 6 new covering empty input, trailing-run extraction, is_error
propagation, boundary at first non-tool message, no-trailing-tool case,
and the end-to-end path from hook -> Turn.tool_results.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add supported providers to prompt caching doc
* Move Z.ai / GLM to cache_control marker list
* Mark xAI models as supporting prompt caching
* Narrow xAI prompt caching flag to models with documented cache pricing
* Add prompt caching flag to grok-4, grok-4-0709, grok-4-latest
---------
Co-authored-by: Michael Riad Zaky <michaelr@Michaels-MacBook-Air.local>
- Use 'auto_router/adaptive_router' prefix in example yaml, docs, and
README — the old 'adaptive_router/...' and 'openai/gpt-4o-mini' values
silently skipped adaptive-router init because detection requires the
'auto_router/adaptive_router' prefix.
- Read x-litellm-min-quality-tier from request headers (and the
'min_quality_tier' metadata key as fallback) in async_pre_routing_hook.
Previously the documented header was defined but never extracted, so
the quality-floor feature was inert.
- Evict expired entries from _session_states. The cache grew without
bound — added a parallel expiry map (same TTL as _owner_cache) and an
opportunistic bulk sweep when the cache crosses a size threshold.
- Align adaptive-router migration SQL with Prisma schema: all count
columns and the 'clean_credit_awarded' / 'last_processed_turn' fields
are NOT NULL in the data model, so the migration now declares them
NOT NULL. Fixes test_aaaasschema_migration_check.
Tests: 8 new covering header/metadata/precedence/invalid-value paths for
min_quality_tier and TTL-based eviction of _session_states.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor: new agentic loop event hook
simplifies how to create logic for tool based multi llm calls
* fix: compress - make it work on anthropic input as well
* fix(compress.py): working prompt compression for claude code
ensures claude code messages can run through proxy easily
* docs: add agentic loop hook guide
* docs: add agentic_loop_hook to sidebar
* fix: fix multiple arguments error
* fix: fix tool call loop for compression on streaming /v1/messages
* fix: fix linting errors
* fix: fix ci/cd errors
* feat(litellm_pre_call_utils.py): use claude code session for litellm session id
allows claude code logs to be stitched together, making it easy to know they were all part of the same conversation
* fix: suppress incorrect mypy warning rE: module
* revert: drop PR's changes to litellm/proxy/_experimental/out/
Restores the 34 HTML files under _experimental/out/ to their pre-PR
paths (X/index.html -> X.html). All renames are R100 (content
unchanged); no other files are touched.
* fix: address greptile review comments on PR #25729
- Skip ``kwargs["tools"] = []`` injection when compression is a no-op —
Anthropic Messages rejects empty tool arrays on requests that did not
originally declare tools.
- Move agentic-loop safety guards (fingerprint cycle / max depth) out of
the per-callback try/except so they propagate instead of being swallowed
by the generic exception handler. Extracted _check_agentic_loop_safety.
- Gate generic ``x-<vendor>-session-id`` capture behind the
LITELLM_CAPTURE_VENDOR_SESSION_HEADERS env var (off by default) to
preserve backwards compatibility; explicit x-litellm-* headers are
unaffected.
- Fix monkeypatch target in pre-call-hook test to patch the actual
module-level binding
(litellm.integrations.compression_interception.handler.compress).
- Add regression tests for empty-tools skip and opt-in session capture.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* revert: drop LITELLM_CAPTURE_VENDOR_SESSION_HEADERS flag
Generic x-<vendor>-session-id header capture is a new feature and only
runs *after* the explicit x-litellm-trace-id / x-litellm-session-id
checks, so it does not change behavior for any existing caller that was
already using the LiteLLM headers — no backwards-incompatibility to gate.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(compress): replace input_type with CallTypes call_type
Drop the bespoke ``CompressionInputType`` literal and use the existing
``litellm.types.utils.CallTypes`` enum instead. ``litellm.compress()``
now takes ``call_type: Union[CallTypes, str]`` (default
``CallTypes.completion``) — no new concept to learn, and the enum is
already the way the rest of the codebase talks about request shapes.
Supported values: ``completion`` / ``acompletion`` (OpenAI chat-completions
shape) and ``anthropic_messages`` (Anthropic structured content blocks).
Updated: compress(), the compression_interception handler, tests, docs,
and the two eval scripts.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- Use LoggingCallbackManager.add_litellm_callback instead of
litellm.callbacks.append (required by callback_manager_test)
- init_adaptive_router_deployment now uses model_name_to_deployment_indices
for O(k) lookup instead of scanning model_list
- Rephrase comment in set_model_list to avoid the 'in self.model_list'
substring that the linear-scan test greps for
- Whitelist _finalize_adaptive_router_if_configured in
test_no_linear_scans_in_router — prefix match on 'auto_router/adaptive_router'
has no supporting index; runs once at init
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Router coverage check flagged this method as untested. Adds two cases:
- initializes AdaptiveRouter from model_list and is idempotent on re-entry
- no-op when no adaptive deployments are configured
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Unrelated timestamp and version drift was showing in the PR diff. This PR adds no new deps — keep uv.lock identical to main.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Remove _experimental/out/ changes from this PR — these are auto-generated Next.js build outputs, not part of the adaptive router feature.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* bump litellm-proxy-extras version to 0.4.67
* bump litellm-proxy-extras pin to 0.4.67 in litellm pyproject
* regenerate uv.lock for litellm-proxy-extras 0.4.67
* bump litellm-enterprise version to 0.1.38
* bump litellm-enterprise pin to 0.1.38 in litellm pyproject
* regenerate uv.lock for litellm-enterprise 0.1.38
send_max_budget_alert_email previously guarded with `is not None`, which
accepts `[]` and then crashes on `recipient_emails[0]` inside
_get_email_params. The current caller (_handle_multi_threshold_max_budget_alert)
already filters empty lists upstream, but the public method signature makes
no such guarantee — a future caller passing [] would hit IndexError.
Switch to truthiness so both None and [] fall through to the single-recipient
path.
Per-user OAuth MCP requests now only skip pre-emptive 401 when a stored token is available, preserving token-reuse behavior while restoring fast PKCE kickoff for first-time or missing-token users.
The lazy import from litellm_enterprise inside _normalize_alert_emails
coupled the core proxy auth path to an optional package. Core should not
depend on enterprise, even lazily — it hides the dependency from static
analysis and inverts the intended layering.
Duplicate the 7-line parser locally. It's pure and unlikely to drift; the
enterprise copy stays where it is for its own callers.
_merge_budget_alert_email_configs previously called list() directly on each
threshold's value, which raised TypeError on null YAML values and silently
split bare strings into single characters. Both are reachable from user-
supplied global config and per-key metadata, so the crash could fire on
every authenticated request once the metadata was in place.
Route both inputs through a _normalize_alert_emails helper that delegates
to the existing _parse_email_list parser (lazy-imported, matching the
enterprise import pattern used elsewhere in proxy/). The merge body keeps
its tight Dict[str, List[str]] contract.
The Union[str, List[str]] value type was speculative — _merge_budget_alert_email_configs
always returns List[str] values, and no caller produces bare strings. Narrowing to
match the runtime guarantee resolves a mypy invariance error at auth_checks.py:3021
without adding casts or Mapping covariance.
A previous refactor added `litellm/integrations/prometheus_helpers.py` as a
sibling to the existing `litellm/integrations/prometheus_helpers/` directory
(which contains `prometheus_api.py` and has no `__init__.py`). The file
shadowed the namespace-package directory, so any deferred
`from litellm.integrations.prometheus_helpers.prometheus_api import ...`
raised `ModuleNotFoundError: 'litellm.integrations.prometheus_helpers' is
not a package` at request time.
Two runtime call sites hit that path:
- /global/spend/logs (spend_management_endpoints.py) returned plain-text 500
"Internal Server Error" for every call, breaking the Admin UI Usage tab
and programmatic consumers.
- SlackAlerting.send_fallback_stats_from_prometheus silently failed inside
its own try/except.
Fix: move prometheus_helpers.py content into prometheus_helpers/__init__.py
and delete the stray .py. The directory becomes a regular package, so both
the package-root import (from ...prometheus_helpers import X) and the
submodule import (from ...prometheus_helpers.prometheus_api import X)
resolve correctly. No call sites change.
Six CI jobs create a miniconda env with python=3.9 before installing
the project; these jobs now fail resolution because the project
requires-python is >=3.10. Bump the conda env python to 3.10 to match
the new floor.
All three dependency bumps in this PR resolve on Python 3.10, so there
is no need to jump the floor all the way to 3.11. Also restore the
py3.10-specific lunary==1.4.36 pin that was collapsed when the floor
was temporarily at 3.11.
Now that requires-python starts at 3.11, the "python_version >= '3.9'"
and ">= '3.10'" markers are unconditionally true, and the "< '3.10'"
entries for psycopg, Pillow, pyarrow, langchain, lunary, and pylint can
never resolve. Drop the dead markers and remove the unreachable pins so
the dependency list reflects what actually gets installed.
Premium fields like policies are echoed at the top level of the
/key/update response, not necessarily mirrored into metadata. Read
metadata first then fall back to the top-level property so an
intentional clear is preserved in either shape.
The /key/update response echoes top-level defaults like policies:[] into
client state. On a subsequent edit, the form resends policies:[], which
the backend treats as "user is setting policies" and blocks with a 403
enterprise check regardless of value.
Drop premium metadata fields from the update payload when the current
form value and the previously persisted value are both empty. Genuine
clears (non-empty -> empty) still pass through so premium users can
clear policies as intended.
Bumps orjson, fastapi-sso, and python-multipart to their latest releases
in the proxy extra, and raises the project python floor to 3.11 so the
updated pins can resolve. CI already runs on 3.11 / 3.12 / 3.13 and the
Docker images ship python 3.13, so the floor change aligns the declared
support range with what is actually tested and shipped.
