scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-05 08:07:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 137

Merge branch 'litellm_internal_staging' of https://github.com/BerriAI/litellm into friendli/glm-5.3

Browse source 
# Conflicts:
#	litellm/model_prices_and_context_window_backup.json
#	model_prices_and_context_window.json
This commit is contained in:
mateo-berri 2026-08-31 15:11:06 -07:00
commit abbccd3fd6
106 changed files with 5846 additions and 1549 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

230
.github/scripts/close_duplicate_issues.py vendored
View file

@ -1,230 +0,0 @@
#!/usr/bin/env python3
"""
Detect and close duplicate GitHub issues using title similarity.
Modes:
--scan Compare all open issues against each other (batch)
--issue-number N Check a single issue against older open issues
Requires the `gh` CLI to be authenticated.
"""
import argparse
import difflib
import json
import re
import subprocess
import sys
def normalize_title(title: str) -> str:
"""Strip common prefixes, lowercase, and collapse whitespace."""
title = re.sub(
r"^\[?(bug|feature request|enhancement|question|docs)[:\]]?\s*",
"",
title,
flags=re.IGNORECASE,
)
return " ".join(title.lower().split())
def gh(*args: str) -> str:
"""Run a gh CLI command and return stdout."""
result = subprocess.run(
["gh", *args],
capture_output=True,
text=True,
check=True,
)
return result.stdout
def fetch_open_issues(repo: str | None) -> list[dict]:
"""Fetch all open issues (excluding PRs) via gh api --paginate."""
if repo:
endpoint = (
f"repos/{repo}/issues?state=open&per_page=100&sort=created&direction=asc"
)
else:
endpoint = "repos/{owner}/{repo}/issues?state=open&per_page=100&sort=created&direction=asc"
cmd = ["api", "--paginate", endpoint]
raw = gh(*cmd)
# gh --paginate concatenates JSON arrays, so we may get multiple arrays
issues = []
for line in raw.strip().splitlines():
line = line.strip()
if not line:
continue
parsed = json.loads(line)
if isinstance(parsed, list):
issues.extend(parsed)
else:
issues.append(parsed)
# Filter out pull requests (they also appear in the issues endpoint)
return [i for i in issues if "pull_request" not in i]
def close_as_duplicate(
issue_number: int, duplicate_of: int, repo: str | None, dry_run: bool
) -> None:
"""Close an issue as duplicate of another, adding a comment and label."""
repo_args = ["--repo", repo] if repo else []
if dry_run:
print(
f" [DRY RUN] Would close #{issue_number} as duplicate of #{duplicate_of}"
)
return
# Add comment
comment_body = (
f"Closing as duplicate of #{duplicate_of}.\n\n"
"If you believe this is not a duplicate, please reopen and add context "
"explaining how this differs."
)
gh("issue", "comment", str(issue_number), "--body", comment_body, *repo_args)
# Add label
gh("issue", "edit", str(issue_number), "--add-label", "duplicate", *repo_args)
# Close with not_planned reason
gh(
"api",
f"repos/{repo or '{owner}/{repo}'}/issues/{issue_number}",
"-X",
"PATCH",
"-f",
"state=closed",
"-f",
"state_reason=not_planned",
)
print(f" Closed #{issue_number} as duplicate of #{duplicate_of}")
def find_duplicate(
issue: dict, candidates: list[dict], threshold: float
) -> dict | None:
"""Return the first candidate whose normalized title is above threshold."""
norm = normalize_title(issue["title"])
for candidate in candidates:
if candidate["number"] == issue["number"]:
continue
cand_norm = normalize_title(candidate["title"])
ratio = difflib.SequenceMatcher(None, norm, cand_norm).ratio()
if ratio >= threshold:
return candidate
return None
def scan_all(
issues: list[dict], threshold: float, repo: str | None, dry_run: bool
) -> int:
"""Compare every issue against all older issues. Returns count of duplicates found."""
# Sort oldest first
issues.sort(key=lambda i: i["number"])
closed_count = 0
for idx, issue in enumerate(issues):
older = issues[:idx]
if not older:
continue
dup = find_duplicate(issue, older, threshold)
if dup:
ratio = difflib.SequenceMatcher(
None,
normalize_title(issue["title"]),
normalize_title(dup["title"]),
).ratio()
print(
f"#{issue['number']}: \"{issue['title']}\"\n"
f" -> duplicate of #{dup['number']}: \"{dup['title']}\" "
f"({ratio:.0%} similar)"
)
close_as_duplicate(issue["number"], dup["number"], repo, dry_run)
closed_count += 1
return closed_count
def check_single(
issue_number: int,
issues: list[dict],
threshold: float,
repo: str | None,
dry_run: bool,
) -> bool:
"""Check a single issue against all older open issues. Returns True if duplicate found."""
target = None
for i in issues:
if i["number"] == issue_number:
target = i
break
if target is None:
print(f"Issue #{issue_number} not found among open issues.")
return False
older = [i for i in issues if i["number"] < issue_number]
dup = find_duplicate(target, older, threshold)
if dup:
ratio = difflib.SequenceMatcher(
None,
normalize_title(target["title"]),
normalize_title(dup["title"]),
).ratio()
print(
f"#{target['number']}: \"{target['title']}\"\n"
f" -> duplicate of #{dup['number']}: \"{dup['title']}\" "
f"({ratio:.0%} similar)"
)
close_as_duplicate(issue_number, dup["number"], repo, dry_run)
return True
print(f"#{issue_number}: no duplicate found above threshold {threshold}")
return False
def main() -> None:
parser = argparse.ArgumentParser(
description="Detect and close duplicate GitHub issues"
)
mode = parser.add_mutually_exclusive_group(required=True)
mode.add_argument("--scan", action="store_true", help="Scan all open issues")
mode.add_argument("--issue-number", type=int, help="Check a single issue number")
parser.add_argument(
"--threshold", type=float, default=0.85, help="Similarity threshold (0-1)"
)
parser.add_argument(
"--close",
action="store_true",
help="Actually close duplicates (default is dry-run)",
)
parser.add_argument(
"--repo", type=str, help="Repository (owner/repo). Auto-detected if omitted."
)
args = parser.parse_args()
dry_run = not args.close
if dry_run:
print("=== DRY RUN MODE (pass --close to actually close issues) ===\n")
print("Fetching open issues...")
issues = fetch_open_issues(args.repo)
print(f"Found {len(issues)} open issues.\n")
if args.scan:
count = scan_all(issues, args.threshold, args.repo, dry_run)
print(f"\nTotal duplicates {'found' if dry_run else 'closed'}: {count}")
else:
found = check_single(
args.issue_number, issues, args.threshold, args.repo, dry_run
)
sys.exit(0 if found else 0) # Always exit 0; finding no dup is not an error
if __name__ == "__main__":
main()

69
.github/workflows/auto-close-duplicates.yml vendored Normal file
View file

@ -0,0 +1,69 @@
name: Auto-close duplicate issues
on:
schedule:
- cron: "0 9 * * *"
workflow_dispatch:
inputs:
dry_run:
description: Log which issues would close without closing anything
type: boolean
default: true
grace_period_days:
description: Days a duplicate notice must go unanswered before the close
type: number
default: 3
pull_request:
paths:
- .github/workflows/auto-close-duplicates.yml
- scripts/auto-close-duplicates.ts
- scripts/auto-close-duplicates.test.ts
permissions: {}
jobs:
test:
if: github.event_name == 'pull_request'
runs-on: ubuntu-latest
timeout-minutes: 5
permissions:
contents: read
steps:
- name: Checkout repository
uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- name: Setup Bun
uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
with:
bun-version: "1.4.0"
- name: Test the sweep
run: bun test scripts/auto-close-duplicates.test.ts
sweep:
if: github.event_name != 'pull_request' && github.repository == 'BerriAI/litellm'
runs-on: ubuntu-latest
timeout-minutes: 10
permissions:
contents: read
issues: write
steps:
- name: Checkout repository
uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- name: Setup Bun
uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
with:
# Exact version, never latest: the next step holds an issues: write token
bun-version: "1.4.0"
- name: Close unanswered duplicates, reopen ones the reporter answered
run: bun run scripts/auto-close-duplicates.ts
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
DRY_RUN: ${{ inputs.dry_run == true }}
GRACE_PERIOD_DAYS: ${{ inputs.grace_period_days }}

40
.github/workflows/check_duplicate_issues.yml vendored
View file

@ -1,12 +1,19 @@
name: Check Duplicate Issues
# Flagging only. "Auto-close duplicate issues" closes a flagged issue 3 days later,
# and only when its title is identical to an older open issue and nobody replied.
# The HTML marker below is the handshake between the two, so keep it in the template.
on:
issues:
types: [opened, edited]
permissions: {}
jobs:
check-duplicate:
runs-on: ubuntu-latest
timeout-minutes: 5
permissions:
issues: write
contents: read
@ -19,35 +26,12 @@ jobs:
threshold: 0.6
reaction: eyes
comment: |
**⚠️ Potential duplicate detected**
<!-- litellm:potential-duplicate candidates={{#issues}}{{number}},{{/issues}} -->
**Potential duplicate detected**
This issue appears similar to existing issue(s):
This looks similar to:
{{#issues}}
- [#{{number}}]({{html_url}}) - {{title}} ({{accuracy}}% similar)
- #{{number}} - {{title}}
{{/issues}}
Please review the linked issue(s) to see if they address your concern. If this is not a duplicate, please provide additional context to help us understand the difference.
- name: Checkout close script
if: github.event.action == 'opened'
uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
sparse-checkout: .github/scripts
persist-credentials: false
- name: Set up Python
if: github.event.action == 'opened'
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: "3.12"
- name: Auto-close if high-confidence duplicate
if: github.event.action == 'opened'
env:
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
run: |
python3 .github/scripts/close_duplicate_issues.py \
--issue-number ${{ github.event.issue.number }} \
--repo ${{ github.repository }} \
--threshold 0.85 \
--close
If this is a duplicate, add a thumbs-up reaction to the existing issue and follow along there. When the title is identical to an older open issue, this issue closes automatically in 3 days unless someone responds. If it is not a duplicate, comment here or add a thumbs-down reaction to this comment and it stays open.

2
CLAUDE.md
View file

@ -23,6 +23,8 @@ When adding new features, add meaningful tests. Don't add tests that don't check
Same thing for bug fixes. The tests should make it so that this specific bug can never happen again without failing tests (i.e., regression)
Never test structure of code only function of it
`tests/test_litellm/` mirrors `litellm/` in a parallel path (see `tests/test_litellm/readme.md`). Name tests `test_<filename>.py`, but always match the existing test file in the directory you touch — many provider dirs use longer descriptive names (e.g. `test_anthropic_chat_transformation.py`) to avoid ambiguity across sibling folders. For bug fixes, extend the existing mapped test file rather than creating a new one. Only create a new test file for a new feature (provider, endpoint, or transformation module) that has no mapped test yet, following that directory's naming convention (or `test_<filename>.py` if you're the first test there). One focused regression test beats many shallow ones
End-to-end tests belong in `tests/e2e/` and must follow the harness conventions documented in that directory's `CLAUDE.md`

15
Dockerfile
View file

@ -1,10 +1,10 @@
# syntax=docker/dockerfile:1.7
# Base image for building
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
# Runtime image
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@ -40,8 +40,8 @@ COPY --from=uvbin /uvx /usr/local/bin/uvx
RUN apk add --no-cache \
bash \
gcc \
python3 \
python3-dev \
python-3.13 \
python-3.13-dev \
rust \
openssl \
openssl-dev \
@ -51,6 +51,7 @@ RUN apk add --no-cache \
ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
UV_LINK_MODE=copy \
UV_PYTHON_DOWNLOADS=0 \
PATH="/app/.venv/bin:${PATH}"
# Copy dependency metadata first for layer caching
@ -65,7 +66,7 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--python python3.13
# Copy full source tree
COPY . .
@ -86,7 +87,7 @@ RUN uv sync --frozen --no-default-groups --no-editable \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -101,7 +102,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
# node (without npm) is required by the prisma CLI at runtime
RUN apk add --no-cache bash openssl tzdata nodejs python3 libsndfile
RUN apk add --no-cache bash openssl tzdata nodejs python-3.13 libsndfile
WORKDIR /app
ENV PATH="/app/.venv/bin:${PATH}" \

12
backend/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin
@ -16,7 +16,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
RUN for i in 1 2 3; do \
apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done
@ -46,7 +46,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra proxy-runtime \
--extra extra_proxy \
--extra semantic-router \
--python python3
--python python3.13
# Stage 2 — copy source and install the project + workspace members.
COPY . .
@ -57,7 +57,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra proxy-runtime \
--extra extra_proxy \
--extra semantic-router \
--python python3
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -71,7 +71,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
RUN for i in 1 2 3; do \
apk add --no-cache bash openssl tzdata python3 libsndfile libatomic && break; \
apk add --no-cache bash openssl tzdata python-3.13 libsndfile libatomic && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done

24
basedpyright-code-budget.json
View file

@ -1,9 +1,9 @@
{
"reportAny": {
"limit": 16389
"limit": 16171
},
"reportArgumentType": {
"limit": 2229
"limit": 2224
},
"reportAssignmentType": {
"limit": 319
@ -18,13 +18,13 @@
"limit": 40
},
"reportDeprecated": {
"limit": 212
"limit": 211
},
"reportDuplicateImport": {
"limit": 19
},
"reportExplicitAny": {
"limit": 5242
"limit": 5199
},
"reportFunctionMemberAccess": {
"limit": 7
@ -54,10 +54,10 @@
"limit": 0
},
"reportMissingParameterType": {
"limit": 5614
"limit": 5611
},
"reportMissingTypeArgument": {
"limit": 15356
"limit": 15350
},
"reportMissingTypeStubs": {
"limit": 40
@ -93,25 +93,25 @@
"limit": 181
},
"reportTypedDictNotRequiredAccess": {
"limit": 25
"limit": 24
},
"reportUndefinedVariable": {
"limit": 0
},
"reportUnknownArgumentType": {
"limit": 44389
"limit": 44368
},
"reportUnknownLambdaType": {
"limit": 109
},
"reportUnknownMemberType": {
"limit": 38500
"limit": 38468
},
"reportUnknownParameterType": {
"limit": 19673
"limit": 19665
},
"reportUnknownVariableType": {
"limit": 30092
"limit": 30066
},
"reportUnnecessaryCast": {
"limit": 111
@ -123,7 +123,7 @@
"limit": 5
},
"reportUnnecessaryIsInstance": {
"limit": 829
"limit": 828
},
"reportUntypedBaseClass": {
"limit": 0

15
docker/Dockerfile.database
View file

@ -1,10 +1,10 @@
# syntax=docker/dockerfile:1.7
# Base image for building
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
# Runtime image
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@ -39,8 +39,8 @@ COPY --from=uvbin /uvx /usr/local/bin/uvx
RUN apk add --no-cache \
bash \
gcc \
python3 \
python3-dev \
python-3.13 \
python-3.13-dev \
openssl \
openssl-dev \
nodejs \
@ -49,6 +49,7 @@ RUN apk add --no-cache \
ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
UV_LINK_MODE=copy \
UV_PYTHON_DOWNLOADS=0 \
PATH="/app/.venv/bin:${PATH}"
# Copy dependency metadata first for layer caching
@ -63,7 +64,7 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--python python3.13
# Copy full source tree
COPY . .
@ -84,7 +85,7 @@ RUN uv sync --frozen --no-default-groups --no-editable \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -98,7 +99,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
# node (without npm) is required by the prisma CLI at runtime
RUN apk add --no-cache bash openssl tzdata nodejs python3 libsndfile
RUN apk add --no-cache bash openssl tzdata nodejs python-3.13 libsndfile
WORKDIR /app
ENV PATH="/app/.venv/bin:${PATH}" \

17
docker/Dockerfile.non_root
View file

@ -1,8 +1,8 @@
# syntax=docker/dockerfile:1.7
# Base images
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG PROXY_EXTRAS_SOURCE=published
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
@ -37,8 +37,8 @@ COPY --from=uvbin /uvx /usr/local/bin/uvx
RUN for i in 1 2 3; do \
apk add --no-cache \
python3 \
python3-dev \
python-3.13 \
python-3.13-dev \
gcc \
rust \
bash \
@ -52,6 +52,7 @@ RUN for i in 1 2 3; do \
ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
UV_LINK_MODE=copy \
UV_PYTHON_DOWNLOADS=0 \
PATH="/app/.venv/bin:${PATH}" \
LITELLM_NON_ROOT=true \
XDG_CACHE_HOME=/app/.cache
@ -69,7 +70,7 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--python python3.13
# Copy full source tree
COPY . .
@ -96,7 +97,7 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3 \
--python python3.13 \
--no-sources-package litellm-proxy-extras; \
else \
uv sync --frozen --no-default-groups --no-editable \
@ -105,7 +106,7 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3; \
--python python3.13; \
fi
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
@ -124,7 +125,7 @@ RUN for i in 1 2 3; do \
apk upgrade --no-cache && break || sleep 5; \
done && \
for i in 1 2 3; do \
apk add --no-cache python3 bash openssl tzdata libsndfile nodejs && break || sleep 5; \
apk add --no-cache python-3.13 bash openssl tzdata libsndfile nodejs && break || sleep 5; \
done
# Copy only what runtime needs. The application is installed inside the venv;

12
gateway/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin
@ -16,7 +16,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
RUN for i in 1 2 3; do \
apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done
@ -47,7 +47,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra extra_proxy \
--extra semantic-router \
--extra bedrock-realtime \
--python python3
--python python3.13
# Stage 2 — copy source and install the project + workspace members.
COPY . .
@ -59,7 +59,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra extra_proxy \
--extra semantic-router \
--extra bedrock-realtime \
--python python3
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -73,7 +73,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
RUN for i in 1 2 3; do \
apk add --no-cache bash openssl tzdata python3 libsndfile libatomic && break; \
apk add --no-cache bash openssl tzdata python-3.13 libsndfile libatomic && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done

1
litellm/constants.py
View file

@ -1728,6 +1728,7 @@ SENTRY_DENYLIST: Final = [
"jwt_token",
"private_key",
"SLACK_WEBHOOK_URL",
"ALERTING_WEBHOOK_URL",
"webhook_url",
"LANGFUSE_SECRET_KEY",
# Email Configuration

167
litellm/google_genai/adapters/transformation.py
View file

@ -1,8 +1,9 @@
import json
from collections.abc import AsyncIterator, Iterator, Sequence
from typing import Any, Final, TypedDict, cast
from collections.abc import AsyncIterator, Callable, Iterator, Mapping, Sequence
from types import MappingProxyType
from typing import Any, Final, TypeAlias, cast
from typing_extensions import ReadOnly
from typing_extensions import ReadOnly, TypedDict
from litellm import verbose_logger
from litellm.litellm_core_utils.json_validation_rule import normalize_tool_schema
@ -11,7 +12,6 @@ from litellm.types.llms.openai import (
ChatCompletionAssistantMessage,
ChatCompletionAssistantToolCall,
ChatCompletionImageObject,
ChatCompletionRequest,
ChatCompletionSystemMessage,
ChatCompletionTextObject,
ChatCompletionToolCallFunctionChunk,
@ -23,35 +23,63 @@ from litellm.types.llms.openai import (
from litellm.types.router import GenericLiteLLMParams
from litellm.types.utils import (
AdapterCompletionStreamWrapper,
ChatCompletionDeltaCustomToolCall,
ChatCompletionDeltaToolCall,
ChatCompletionMessageCustomToolCall,
ChatCompletionMessageToolCall,
Choices,
Delta,
Function,
Message,
ModelResponse,
ModelResponseStream,
StreamingChoices,
Usage,
)
class _GenAITextPart(TypedDict, total=False):
text: ReadOnly[str]
_JsonDict: TypeAlias = dict[str, object]
_JsonDictList: TypeAlias = list[_JsonDict]
class _GenAISystemInstruction(TypedDict, total=False):
parts: ReadOnly[list[_GenAITextPart]]
class _ToolCallAccumulator(TypedDict):
name: ReadOnly[str]
arguments: ReadOnly[str]
class _GenAIFunctionCall(TypedDict):
name: ReadOnly[str]
args: ReadOnly[Mapping[str, object]]
class _GenAIPart(TypedDict, total=False):
text: ReadOnly[str]
functionCall: ReadOnly[dict[str, object]]
functionCall: ReadOnly[_GenAIFunctionCall]
class _GenAIFunctionResponse(TypedDict, total=False):
name: ReadOnly[str]
response: ReadOnly[object]
class _GenAIRequestFunctionCall(TypedDict, total=False):
name: ReadOnly[str]
args: ReadOnly[Mapping[str, object]]
class _GenAIContentPart(TypedDict, total=False):
text: ReadOnly[str]
inline_data: ReadOnly[Mapping[str, str]]
functionResponse: ReadOnly[_GenAIFunctionResponse]
functionCall: ReadOnly[_GenAIRequestFunctionCall]
class _GenAIFunctionDeclaration(TypedDict, total=False):
name: ReadOnly[str]
description: ReadOnly[str]
parametersJsonSchema: ReadOnly[dict[str, object]]
parametersJsonSchema: ReadOnly[object]
class _GenAITool(TypedDict, total=False):
functionDeclarations: ReadOnly[list[_GenAIFunctionDeclaration]]
functionDeclarations: ReadOnly[Sequence[_GenAIFunctionDeclaration]]
class _GenAIFunctionCallingConfig(TypedDict, total=False):
@ -62,9 +90,11 @@ class _GenAIToolConfig(TypedDict, total=False):
functionCallingConfig: ReadOnly[_GenAIFunctionCallingConfig]
def _decode_tool_call_arguments(raw_arguments: str) -> object:
"""Decode a tool call's JSON-encoded arguments into the value Google GenAI expects."""
return json.loads(raw_arguments)
class _GenAISystemInstruction(TypedDict, total=False):
parts: ReadOnly[Sequence[Mapping[str, str]]]
_EMPTY_STR_MAPPING: Final[Mapping[str, str]] = MappingProxyType({})
class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
@ -74,12 +104,12 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
"""
sent_first_chunk: bool = False
# State tracking for accumulating partial tool calls
accumulated_tool_calls: dict[int, dict[str, str]]
_parse_accumulated_args: Callable[[str], Mapping[str, object]] = staticmethod(json.loads)
def __init__(self, completion_stream: object):
self.sent_first_chunk = False
self.accumulated_tool_calls = {}
# State tracking for accumulating partial tool calls
self.accumulated_tool_calls = dict[int, _ToolCallAccumulator]()
self._returned_response = False
super().__init__(completion_stream)
@ -124,7 +154,7 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
# After the stream is exhausted, check for any remaining accumulated tool calls
if self.accumulated_tool_calls:
try:
parts: Final[list[_GenAIPart]] = []
parts: Final = list[_GenAIPart]()
for (
tool_call_index,
tool_call_data,
@ -132,7 +162,9 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
try:
# For tool calls with no arguments, accumulated_args will be "", which is not valid JSON.
# We default to an empty JSON object in this case.
parsed_args = _decode_tool_call_arguments(tool_call_data["arguments"] or "{}")
parsed_args: Mapping[str, object] = self._parse_accumulated_args(
tool_call_data["arguments"] or "{}"
)
function_call_part: _GenAIPart = {
"functionCall": {
"name": tool_call_data["name"] or "undefined_tool_name",
@ -149,7 +181,7 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
tool_call_data["arguments"],
)
if parts:
final_chunk: Final[dict[str, object]] = {
final_chunk: Final = {
"candidates": [
{
"content": {"parts": parts, "role": "model"},
@ -211,14 +243,16 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
class GoogleGenAIAdapter:
"""Adapter for transforming Google GenAI generate_content requests to/from litellm.completion format"""
_parse_tool_call_args: Callable[[str], Mapping[str, object]] = staticmethod(json.loads)
def __init__(self) -> None:
pass
def translate_generate_content_to_completion(
self,
model: str,
contents: list[dict[str, Any]] | dict[str, Any],
config: dict[str, Any] | None = None,
contents: _JsonDictList | _JsonDict,
config: Mapping[str, object] | None = None,
litellm_params: GenericLiteLLMParams | None = None,
**kwargs,
) -> dict[str, Any]:
@ -250,7 +284,7 @@ class GoogleGenAIAdapter:
messages: Final = self._transform_contents_to_messages(contents_list, system_instruction=system_instruction)
# Create base request as dict (which is compatible with ChatCompletionRequest)
completion_request: Final[ChatCompletionRequest] = {
completion_request: Final[_JsonDict] = {
"model": model,
"messages": messages,
}
@ -312,9 +346,9 @@ class GoogleGenAIAdapter:
def _add_generic_litellm_params_to_request(
self,
completion_request_dict: dict[str, object],
completion_request_dict: _JsonDict,
litellm_params: GenericLiteLLMParams | None = None,
) -> dict[str, object]:
) -> _JsonDict:
"""Add generic litellm params to request. e.g add api_base, api_key, api_version, etc.
Args:
@ -326,7 +360,7 @@ class GoogleGenAIAdapter:
"""
allowed_fields: Final = GenericLiteLLMParams.model_fields.keys()
if litellm_params:
litellm_dict: Final = litellm_params.model_dump(exclude_none=True)
litellm_dict: Final[_JsonDict] = litellm_params.model_dump(exclude_none=True)
for key, value in litellm_dict.items():
if key in allowed_fields:
completion_request_dict[key] = value
@ -346,12 +380,12 @@ class GoogleGenAIAdapter:
tools: Sequence[_GenAITool],
) -> list[ChatCompletionToolParam]:
"""Transform Google GenAI tools to OpenAI tools format"""
openai_tools: Final[list[dict[str, object]]] = []
openai_tools: Final = list[_JsonDict]()
for tool in tools:
if "functionDeclarations" in tool:
for func_decl in tool["functionDeclarations"]:
function_chunk: dict[str, object] = {
function_chunk: _JsonDict = {
"name": func_decl.get("name", ""),
}
@ -360,7 +394,7 @@ class GoogleGenAIAdapter:
if "parametersJsonSchema" in func_decl:
function_chunk["parameters"] = func_decl["parametersJsonSchema"]
openai_tool: dict[str, object] = {"type": "function", "function": function_chunk}
openai_tool: _JsonDict = {"type": "function", "function": function_chunk}
openai_tools.append(openai_tool)
# normalize the tool schemas
@ -391,13 +425,13 @@ class GoogleGenAIAdapter:
# Handle system instruction
if system_instruction:
system_parts: Final = system_instruction.get("parts", [])
system_parts: Final[Sequence[Mapping[str, str]]] = system_instruction.get("parts", [])
if system_parts and "text" in system_parts[0]:
messages.append(ChatCompletionSystemMessage(role="system", content=system_parts[0]["text"]))
for content in contents:
role = content.get("role", "user")
parts = content.get("parts", [])
parts: Sequence[_GenAIContentPart | str | None] = content.get("parts", [])
if role == "user":
# Handle user messages with potential function responses
@ -500,7 +534,7 @@ class GoogleGenAIAdapter:
def translate_completion_to_generate_content(
self,
response: ModelResponse,
) -> dict[str, object]:
) -> _JsonDict:
"""
Transform litellm completion response to Google GenAI generate_content format
@ -523,13 +557,13 @@ class GoogleGenAIAdapter:
parts = self._transform_openai_message_to_google_genai_parts(choice.message)
else:
# Fallback for generic choice objects
message_content = getattr(choice, "message", {}).get("content", "") or getattr(choice, "delta", {}).get(
"content", ""
)
message_content: str = getattr(choice, "message", _EMPTY_STR_MAPPING).get("content", "") or getattr(
choice, "delta", _EMPTY_STR_MAPPING
).get("content", "")
parts = [{"text": message_content}] if message_content else []
# Create Google GenAI format response
generate_content_response: Final[dict[str, object]] = {
generate_content_response: Final[_JsonDict] = {
"candidates": [
{
"content": {"parts": parts, "role": "model"},
@ -563,7 +597,7 @@ class GoogleGenAIAdapter:
self,
response: ModelResponse | ModelResponseStream,
wrapper: GoogleGenAIStreamWrapper,
) -> dict[str, object] | None:
) -> Mapping[str, object] | None:
"""
Transform streaming litellm completion chunk to Google GenAI generate_content format
@ -590,7 +624,7 @@ class GoogleGenAIAdapter:
finish_reason: str | None = getattr(choice, "finish_reason", None)
else:
# Fallback for generic choice objects
message_content: Final = getattr(choice, "delta", {}).get("content", "")
message_content: Final[str] = getattr(choice, "delta", _EMPTY_STR_MAPPING).get("content", "")
parts = [{"text": message_content}] if message_content else []
finish_reason = getattr(choice, "finish_reason", None)
@ -599,7 +633,7 @@ class GoogleGenAIAdapter:
return None
# Create Google GenAI streaming format response
streaming_chunk: Final[dict[str, object]] = {
streaming_chunk: Final[_JsonDict] = {
"candidates": [
{
"content": {"parts": parts, "role": "model"},
@ -635,10 +669,10 @@ class GoogleGenAIAdapter:
def _transform_openai_message_to_google_genai_parts(
self,
message: Any,
) -> list[_GenAIPart]:
message: Message,
) -> Sequence[_GenAIPart]:
"""Transform OpenAI message to Google GenAI parts format"""
parts: Final[list[_GenAIPart]] = []
parts: Final = list[_GenAIPart]()
# Add text content if present
if hasattr(message, "content") and message.content:
@ -646,20 +680,22 @@ class GoogleGenAIAdapter:
# Add tool calls if present
if hasattr(message, "tool_calls") and message.tool_calls:
for tool_call in message.tool_calls:
if hasattr(tool_call, "function") and tool_call.function:
tool_calls: Final[Sequence[ChatCompletionMessageToolCall | ChatCompletionMessageCustomToolCall]] = (
message.tool_calls
)
for tool_call in tool_calls:
function: Function | None = getattr(tool_call, "function", None)
if function:
try:
args = (
_decode_tool_call_arguments(tool_call.function.arguments)
if tool_call.function.arguments
else {}
args: Mapping[str, object] = (
self._parse_tool_call_args(function.arguments) if function.arguments else {}
)
except json.JSONDecodeError:
args = {}
function_call_part: _GenAIPart = {
"functionCall": {
"name": tool_call.function.name or "undefined_tool_name",
"name": function.name or "undefined_tool_name",
"args": args,
}
}
@ -668,21 +704,23 @@ class GoogleGenAIAdapter:
return parts if parts else [{"text": ""}]
def _transform_openai_delta_to_google_genai_parts_with_accumulation(
self, delta: Any, wrapper: GoogleGenAIStreamWrapper
) -> list[_GenAIPart]:
self, delta: Delta, wrapper: GoogleGenAIStreamWrapper
) -> Sequence[_GenAIPart]:
"""Transforms OpenAI delta to Google GenAI parts, accumulating streaming tool calls."""
# 1. Initialize wrapper state if it doesn't exist
if not hasattr(wrapper, "accumulated_tool_calls"):
wrapper.accumulated_tool_calls = {}
parts: Final[list[_GenAIPart]] = []
parts: Final = list[_GenAIPart]()
if hasattr(delta, "content") and delta.content:
parts.append({"text": delta.content})
# 2. Ensure tool_calls is iterable
tool_calls: Final = delta.tool_calls or []
tool_calls: Final[Sequence[ChatCompletionDeltaToolCall | ChatCompletionDeltaCustomToolCall]] = (
delta.tool_calls or []
)
for tool_call in tool_calls:
if not hasattr(tool_call, "function"):
@ -701,19 +739,20 @@ class GoogleGenAIAdapter:
}
# Accumulate name and arguments
function_name = getattr(tool_call.function, "name", None)
args_chunk = getattr(tool_call.function, "arguments", None)
delta_function: Function | None = getattr(tool_call, "function", None)
function_name: str | None = getattr(delta_function, "name", None)
args_chunk: str | None = getattr(delta_function, "arguments", None)
# Optimization: Skip chunks that have no new data
if not function_name and not args_chunk:
verbose_logger.debug("Skipping empty tool call chunk for index: %s", tool_call_index)
continue
if function_name:
wrapper.accumulated_tool_calls[tool_call_index]["name"] = function_name
if args_chunk:
wrapper.accumulated_tool_calls[tool_call_index]["arguments"] += args_chunk
previous_data: _ToolCallAccumulator = wrapper.accumulated_tool_calls[tool_call_index]
wrapper.accumulated_tool_calls[tool_call_index] = _ToolCallAccumulator(
name=function_name or previous_data["name"],
arguments=previous_data["arguments"] + (args_chunk or ""),
)
# Attempt to parse and emit a complete tool call
accumulated_data = wrapper.accumulated_tool_calls[tool_call_index]
@ -723,7 +762,7 @@ class GoogleGenAIAdapter:
# 5. Attempt to parse arguments even if name hasn't arrived.
try:
# Attempt to parse the accumulated arguments string
parsed_args = _decode_tool_call_arguments(accumulated_args)
parsed_args: Mapping[str, object] = self._parse_tool_call_args(accumulated_args)
# If parsing succeeds, but we don't have a name yet, wait.
# The part will be created by a later chunk that brings the name.
@ -757,7 +796,7 @@ class GoogleGenAIAdapter:
return mapping.get(finish_reason, "STOP")
def _map_usage(self, usage: Usage | None) -> dict[str, int]:
def _map_usage(self, usage: object) -> Mapping[str, int]:
"""Map OpenAI usage to Google GenAI usage format"""
return {
"promptTokenCount": getattr(usage, "prompt_tokens", 0) or 0,

8
litellm/integrations/SlackAlerting/slack_alerting.py
View file

@ -1485,9 +1485,9 @@ Model Info:
elif self.default_webhook_url is not None:
_digest_webhook = self.default_webhook_url
else:
_digest_webhook = os.getenv("SLACK_WEBHOOK_URL", None)
_digest_webhook = os.getenv("SLACK_WEBHOOK_URL") or os.getenv("ALERTING_WEBHOOK_URL")
if _digest_webhook is None:
raise ValueError("Missing SLACK_WEBHOOK_URL from environment")
raise ValueError("Missing SLACK_WEBHOOK_URL / ALERTING_WEBHOOK_URL from environment")
digest_key: Final = f"{alert_type_name_str}:{request_model or ''}:{api_base or ''}"
@ -1516,10 +1516,10 @@ Model Info:
elif self.default_webhook_url is not None:
slack_webhook_url = self.default_webhook_url
else:
slack_webhook_url = os.getenv("SLACK_WEBHOOK_URL", None)
slack_webhook_url = os.getenv("SLACK_WEBHOOK_URL") or os.getenv("ALERTING_WEBHOOK_URL")
if slack_webhook_url is None:
raise ValueError("Missing SLACK_WEBHOOK_URL from environment")
raise ValueError("Missing SLACK_WEBHOOK_URL / ALERTING_WEBHOOK_URL from environment")
payload: Final = {"text": formatted_message}
headers: Final = {"Content-type": "application/json"}

2
litellm/integrations/otel/mappers/genai.py
View file

@ -62,6 +62,8 @@ class GenAIMapper:
GenAI.RESPONSE_TIME_TO_FIRST_CHUNK: lambda d: d.time_to_first_chunk_seconds,
GenAI.USAGE_INPUT_TOKENS: lambda d: d.usage.input_tokens,
GenAI.USAGE_OUTPUT_TOKENS: lambda d: d.usage.output_tokens,
GenAI.USAGE_CACHE_CREATION_INPUT_TOKENS: lambda d: d.usage.cache_creation_input_tokens,
GenAI.USAGE_CACHE_READ_INPUT_TOKENS: lambda d: d.usage.cache_read_input_tokens,
Error.TYPE: lambda d: d.error.error_type if d.error else None,
Server.ADDRESS: lambda d: d.server.address if d.server else None,
Server.PORT: lambda d: d.server.port if d.server else None,

22
litellm/integrations/otel/model/payloads.py
View file

@ -95,6 +95,22 @@ class LLMUsage:
input_tokens: int | None = None
output_tokens: int | None = None
total_tokens: int | None = None
cache_creation_input_tokens: int | None = None
cache_read_input_tokens: int | None = None
@classmethod
def from_standard_logging_payload(cls, payload: StandardLoggingPayload) -> LLMUsage:
# Cache token counts only exist on the raw provider usage object under metadata
metadata: Final[Mapping[str, object]] = payload.get("metadata") or {}
raw_usage: Final = metadata.get("usage_object")
usage_object: Final[Mapping[str, object]] = raw_usage if isinstance(raw_usage, Mapping) else {}
return cls(
input_tokens=as_int(payload.get("prompt_tokens")),
output_tokens=as_int(payload.get("completion_tokens")),
total_tokens=as_int(payload.get("total_tokens")),
cache_creation_input_tokens=as_int(usage_object.get("cache_creation_input_tokens")),
cache_read_input_tokens=as_int(usage_object.get("cache_read_input_tokens")),
)
@dataclass(frozen=True)
@ -363,11 +379,7 @@ class LLMCallSpanData:
response_model=context.response_model,
response_id=as_str(response.get("id")),
request_params=LLMRequestParams.from_model_parameters(params),
usage=LLMUsage(
input_tokens=as_int(payload.get("prompt_tokens")),
output_tokens=as_int(payload.get("completion_tokens")),
total_tokens=as_int(payload.get("total_tokens")),
),
usage=LLMUsage.from_standard_logging_payload(payload),
finish_reasons=finish_reasons,
error=_parse_error(payload),
response_cost=as_float(payload.get("response_cost")),

2
litellm/integrations/otel/model/semconv.py
View file

@ -110,6 +110,8 @@ class GenAI:
# usage
USAGE_INPUT_TOKENS: Final = "gen_ai.usage.input_tokens"
USAGE_OUTPUT_TOKENS: Final = "gen_ai.usage.output_tokens"
USAGE_CACHE_CREATION_INPUT_TOKENS: Final = "gen_ai.usage.cache_creation.input_tokens"
USAGE_CACHE_READ_INPUT_TOKENS: Final = "gen_ai.usage.cache_read.input_tokens"
# content (opt-in, gated by capture mode)
INPUT_MESSAGES: Final = "gen_ai.input.messages"
OUTPUT_MESSAGES: Final = "gen_ai.output.messages"

211
litellm/integrations/rubrik.py
View file

@ -8,11 +8,12 @@ import uuid
from collections import Counter
from collections.abc import Awaitable, Mapping, Sequence
from dataclasses import dataclass
from datetime import datetime
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypedDict
from typing import TYPE_CHECKING, Final, Literal, Optional, Protocol, TypedDict, overload
import httpx
from typing_extensions import Never, ReadOnly
from typing_extensions import Never, ReadOnly, Required
from litellm._logging import verbose_logger
from litellm.integrations.custom_batch_logger import CustomBatchLogger
@ -30,6 +31,7 @@ from litellm.llms.custom_httpx.http_handler import (
httpxSpecialProvider,
)
from litellm.types.guardrails import GuardrailEventHooks
from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolCallChunk
from litellm.types.utils import (
ChatCompletionMessageToolCall,
Function,
@ -52,17 +54,102 @@ _DROP_WARNING_INTERVAL_SECONDS: Final = 60.0
_EMPTY_MAPPING: Final[Mapping[str, Never]] = MappingProxyType({})
class _ServiceToolCall(TypedDict):
id: ReadOnly[str]
class _ModerationToolCall(TypedDict, total=False):
id: ReadOnly[Required[str]]
class _ServiceMessage(TypedDict, total=False):
class _ModerationMessage(TypedDict, total=False):
content: ReadOnly[str | None]
tool_calls: ReadOnly[Sequence[_ModerationToolCall] | None]
class _ModerationChoice(TypedDict, total=False):
message: ReadOnly[_ModerationMessage | None]
class _ModerationResponse(TypedDict, total=False):
choices: ReadOnly[Sequence[_ModerationChoice]]
class _LogEventKwargs(TypedDict, total=False):
standard_logging_object: ReadOnly[Required[StandardLoggingPayload]]
litellm_call_id: ReadOnly[str]
class _HasCallId(Protocol):
def get(self, key: Literal["litellm_call_id"], /) -> str | None: ...
class _HasModelAttr(Protocol):
model: str | None
class _ResponseSource(Protocol):
def get(self, key: Literal["response"], /) -> "_HasModelAttr | None": ...
class _ModelSource(Protocol):
def get(self, key: Literal["model"], default: str, /) -> str: ...
class _FallbackSource(Protocol):
@overload
def get(self, key: Literal["start_time"], /) -> datetime | None: ...
@overload
def get(self, key: str, /) -> object | None: ...
class _RequestContextSource(Protocol):
@overload
def get(self, key: Literal["optional_params"], /) -> Mapping[str, object] | None: ...
@overload
def get(self, key: str, /) -> object | None: ...
def __contains__(self, key: object, /) -> bool: ...
def __getitem__(self, key: str, /) -> object: ...
class _ToolCallLike(Protocol):
id: str | None
type: str | None
function: Function
class _ModerationSourceToolCall(TypedDict, total=False):
function: ReadOnly[Mapping[str, object] | None]
class _ModerationSourceMessage(TypedDict, total=False):
role: ReadOnly[str]
function_call: ReadOnly[Mapping[str, object] | None]
tool_calls: ReadOnly[Sequence[_ModerationSourceToolCall | None] | None]
class _FlattenedModerationMessage(TypedDict):
role: ReadOnly[str | None]
content: ReadOnly[str]
tool_calls: ReadOnly[Sequence[_ServiceToolCall]]
class _ServiceChoice(TypedDict, total=False):
message: ReadOnly[_ServiceMessage]
class _CorrelatablePayload(TypedDict):
id: str # writable-ok: _apply_correlation_id overwrites the provider id on a deep-copied payload
class _SystemPromptCarrier(TypedDict, total=False):
messages: object # writable-ok: _prepend_system_prompt rebinds messages on the copied payload by design
class _BlockFailurePayload(TypedDict, total=False):
id: object # writable-ok: correlation id is pinned after copying the base payload
model: ReadOnly[object]
model_group: ReadOnly[object]
model_id: ReadOnly[str]
model_parameters: ReadOnly[object]
startTime: ReadOnly[float | None]
endTime: ReadOnly[float | None]
completionStartTime: ReadOnly[float | None]
messages: object # writable-ok: passed to _prepend_system_prompt, which rebinds messages
metadata: ReadOnly[StandardLoggingUserAPIKeyMetadata]
response: str # writable-ok: block failure text replaces the copied response
status: ReadOnly[str]
class _MalformedToolBlockingResponseError(Exception):
@ -385,7 +472,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
@staticmethod
def _stash_block_context(
logging_obj: Optional["LiteLLMLoggingObj"],
request_data: dict,
request_data: dict[str, object],
) -> None:
"""Stash signals so the deferred success-event skips this request and
``async_post_call_failure_hook`` can build the failure payload.
@ -414,12 +501,16 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
request_data["_rubrik_logging_obj"] = logging_obj
@staticmethod
def _normalize_tool_calls(tool_calls: Sequence[object]) -> tuple[ChatCompletionMessageToolCall, ...]:
def _normalize_tool_calls(
tool_calls: Sequence[ChatCompletionToolCallChunk | ChatCompletionMessageToolCall | _ToolCallLike],
) -> tuple[ChatCompletionMessageToolCall, ...]:
"""Convert tool_calls from inputs to ChatCompletionMessageToolCall objects."""
return tuple(RubrikLogger._normalize_tool_call(tc) for tc in tool_calls)
@staticmethod
def _normalize_tool_call(tc: Any) -> ChatCompletionMessageToolCall:
def _normalize_tool_call(
tc: ChatCompletionToolCallChunk | ChatCompletionMessageToolCall | _ToolCallLike,
) -> ChatCompletionMessageToolCall:
if isinstance(tc, ChatCompletionMessageToolCall):
return tc
if isinstance(tc, dict):
@ -460,12 +551,15 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
``content`` is sent so the webhook can moderate the response text;
``None`` when the assistant produced no text (tool-call-only response).
"""
message: Final[dict[str, object]] = {
message: Final[Mapping[str, object]] = {
"role": "assistant",
"content": content or None,
**(
{"tool_calls": tuple(tc.model_dump(exclude_none=True) for tc in tool_calls)}
if tool_calls
else _EMPTY_MAPPING
),
}
if tool_calls:
message["tool_calls"] = tuple(tc.model_dump(exclude_none=True) for tc in tool_calls)
return {
"id": request_id or f"chatcmpl-{uuid.uuid4()}",
"object": "chat.completion",
@ -481,7 +575,9 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
}
@staticmethod
def _flatten_messages_for_moderation(messages: Sequence[object] | None) -> tuple[Mapping[str, Any], ...]:
def _flatten_messages_for_moderation(
messages: Sequence[AllMessageValues | None] | None,
) -> tuple[_FlattenedModerationMessage, ...]:
"""Collapse each message's content to a plain string for the webhook.
litellm normalizes Anthropic ``/v1/messages`` requests to OpenAI shape,
@ -502,7 +598,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
)
@staticmethod
def _moderation_text_parts(message: Mapping[str, Any]) -> tuple[str, ...]:
def _moderation_text_parts(message: _ModerationSourceMessage) -> tuple[str, ...]:
"""Every attacker-controlled text segment of a message: its content plus
the arguments of any tool call or deprecated function call."""
fc: Final = message.get("function_call")
@ -530,16 +626,8 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
``/v1/messages`` requests too. Optional fields are sent only when
present so the payload stays clean.
"""
payload: Final[dict[str, object]] = {
"model": inputs.get("model") or request_data.get("model") or "",
"messages": RubrikLogger._flatten_messages_for_moderation(inputs.get("structured_messages")),
}
tools: Final = inputs.get("tools")
if tools is not None:
payload["tools"] = tools
user: Final = request_data.get("user")
if user:
payload["user"] = user
# Fall back to litellm_call_id, the stable cross-provider join key the
# response/tool path uses (see _correlation_id). LiteLLM does not
# populate request_data["correlation_key"]; it carries litellm_call_id.
@ -547,14 +635,18 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
# when correlation_key is empty, so without this the block fires but no
# log is ever written. An explicit correlation_key still wins.
correlation_key: Final = request_data.get("correlation_key") or request_data.get("litellm_call_id")
if correlation_key:
payload["correlation_key"] = correlation_key
return payload
return {
"model": inputs.get("model") or request_data.get("model") or "",
"messages": RubrikLogger._flatten_messages_for_moderation(inputs.get("structured_messages")),
**({"tools": tools} if tools is not None else _EMPTY_MAPPING),
**({"user": user} if user else _EMPTY_MAPPING),
**({"correlation_key": correlation_key} if correlation_key else _EMPTY_MAPPING),
}
@staticmethod
def _extract_request_data(
call_details: Mapping[str, Any],
request_data: Mapping[str, object] | None,
call_details: _RequestContextSource,
request_data: _RequestContextSource | None,
) -> Mapping[str, object]:
"""Extract original request data from model_call_details for the
response moderation service envelope.
@ -590,7 +682,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
}
@staticmethod
def _sanitize_proxy_server_request(proxy_server_request: object) -> object:
def _sanitize_proxy_server_request(proxy_server_request: Mapping[str, object] | str | None) -> object:
"""Allowlist only routing fields (``url``, ``method``) when forwarding
``proxy_server_request`` to an external webhook, dropping inbound
``headers`` (Authorization, Cookie, x-api-key, ...) and the raw
@ -600,18 +692,19 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
return {key: proxy_server_request[key] for key in ("url", "method") if key in proxy_server_request}
@staticmethod
def _resolve_model(request_data: Mapping[str, object], call_details: Mapping[str, str]) -> str:
def _resolve_model(request_data: _ResponseSource, call_details: _ModelSource) -> str:
"""Get the model name for the ModifyResponseException."""
response: Final = request_data.get("response")
if response and hasattr(response, "model"):
response_model: Final[str | None] = getattr(response, "model", None)
return response_model or "unknown"
return response.model or "unknown"
return call_details.get("model", "unknown")
# -- Logging hooks ---------------------------------------------------------
@staticmethod
def _correlation_id(call_details: Mapping[str, str], request_data: Mapping[str, str] | None = None) -> str | None:
def _correlation_id(
call_details: _HasCallId | _LogEventKwargs, request_data: _HasCallId | None = None
) -> str | None:
"""The id that joins a blocked request's two S3 logs by filename: the
moderation (``_blocking``) log and the failure (response) log.
@ -625,7 +718,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
return call_details.get("litellm_call_id") or (request_data or _EMPTY_MAPPING).get("litellm_call_id")
@classmethod
def _apply_correlation_id(cls, payload: dict[str, object], source: Mapping[str, str]) -> None:
def _apply_correlation_id(cls, payload: _CorrelatablePayload, source: _HasCallId | _LogEventKwargs) -> None:
"""Pin ``payload["id"]`` to ``litellm_call_id`` in place so this log
shares its S3 filename id with the moderation (``_blocking``) and
failure logs for the same request -- for every provider.
@ -645,7 +738,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
payload["id"] = correlated
@staticmethod
def _prepend_system_prompt(payload: dict[str, object], source: Mapping[str, object]) -> None:
def _prepend_system_prompt(payload: _SystemPromptCarrier, source: Mapping[str, object]) -> None:
"""Prepend ``source["system"]`` onto ``payload["messages"]``.
Builds a NEW messages list rather than mutating ``payload["messages"]``
@ -673,9 +766,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
exc_info=True,
)
async def _prepare_log_payload(
self, kwargs: Mapping[str, object], event_type: str
) -> StandardLoggingPayload | None:
async def _prepare_log_payload(self, kwargs: _LogEventKwargs, event_type: str) -> StandardLoggingPayload | None:
"""Shared logic for success logging (sampled)."""
if random.random() > self.sampling_rate:
verbose_logger.debug("Skipping Rubrik %s logging (sampling_rate=%s)", event_type, self.sampling_rate)
@ -684,12 +775,12 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
# Deep-copy so mutations don't affect other callbacks sharing this object
standard_logging_payload: Final[StandardLoggingPayload] = safe_deep_copy(kwargs["standard_logging_object"])
self._apply_correlation_id(standard_logging_payload, kwargs) # pyright: ignore[reportArgumentType] # StandardLoggingPayload is dict[str,Any] at runtime
self._apply_correlation_id(standard_logging_payload, kwargs)
self._prepend_system_prompt(standard_logging_payload, kwargs) # pyright: ignore[reportArgumentType] # StandardLoggingPayload is dict[str,Any] at runtime
return standard_logging_payload
async def _append_and_maybe_flush(self, payload) -> None:
async def _append_and_maybe_flush(self, payload: Mapping[str, object]) -> None:
self._ensure_periodic_flush_task()
self.log_queue.append(payload)
self._enforce_max_queue_size()
@ -714,7 +805,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
self._dropped_since_warning = 0
self._last_drop_warning_time = now
async def _enqueue_log_event(self, kwargs: Mapping[str, object], event_type: str):
async def _enqueue_log_event(self, kwargs: _LogEventKwargs, event_type: str):
try:
payload: Final = await self._prepare_log_payload(kwargs, event_type)
if payload is None:
@ -835,7 +926,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
logging_obj: "LiteLLMLoggingObj",
exception: "ModifyResponseException",
user_api_key_dict: "UserAPIKeyAuth",
) -> StandardLoggingPayload:
) -> _BlockFailurePayload:
"""Build a failure-style payload using the exception text as response.
Blocked-tool events are security-relevant and **bypass sampling**:
@ -877,9 +968,9 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
call_details: Final = logging_obj.model_call_details
exception_text: Final = f"{type(exception).__name__}: {exception.message}"
base: Final = call_details.get("standard_logging_object")
base: Final[StandardLoggingPayload | None] = call_details.get("standard_logging_object")
if base is not None:
payload: dict[str, object] = safe_deep_copy(base)
payload: _BlockFailurePayload = self._copy_block_payload_base(base)
else:
verbose_logger.debug(
"Rubrik: standard_logging_object not yet on model_call_details "
@ -901,6 +992,10 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
return payload
@staticmethod
def _copy_block_payload_base(base: StandardLoggingPayload) -> _BlockFailurePayload:
return safe_deep_copy(base)
@staticmethod
def _caller_metadata(user_api_key_dict: "UserAPIKeyAuth") -> StandardLoggingUserAPIKeyMetadata:
"""Identify the caller whose request was blocked.
@ -923,9 +1018,9 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
@classmethod
def _build_fallback_payload(
cls,
call_details: Mapping[str, Any],
call_details: _FallbackSource,
user_api_key_dict: "UserAPIKeyAuth",
) -> dict[str, object]:
) -> _BlockFailurePayload:
# Convert datetime to a Unix float so json.dumps can serialize it.
# httpx's json= parameter uses stdlib json.dumps with no custom encoder.
_raw_start: Final = call_details.get("start_time")
@ -959,7 +1054,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
response: Final = await self.async_httpx_client.post(
url=self.logging_endpoint,
json=data,
headers=self._headers,
headers=dict(self._headers),
)
response.raise_for_status()
except httpx.HTTPStatusError as e:
@ -1013,7 +1108,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
# -- Webhook services ------------------------------------------------------
async def _post_json(self, endpoint: str, payload: Mapping[str, object], service_name: str) -> Mapping[str, Any]:
async def _post_json(self, endpoint: str, payload: Mapping[str, object], service_name: str) -> _ModerationResponse:
"""POST ``payload`` to a Rubrik webhook and return its dict response.
Raises:
@ -1023,11 +1118,11 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
verbose_logger.debug("Sending request to %s: %s", service_name, endpoint)
http_response: Final = await self.moderation_client.post(
endpoint,
json=payload,
headers=self._headers,
json=dict(payload),
headers=dict(self._headers),
)
http_response.raise_for_status()
result: Final[object] = http_response.json()
result: Final[_ModerationResponse | None] = http_response.json()
if not isinstance(result, dict):
raise TypeError(
f"{service_name} returned non-dict JSON "
@ -1040,7 +1135,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
self,
response_data: Mapping[str, object],
request_data: Mapping[str, object],
) -> Mapping[str, Any]:
) -> _ModerationResponse:
"""Post the ``{request, response}`` envelope to the after_completion
webhook and return its (possibly rewritten) response.
@ -1056,7 +1151,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
"Response moderation service",
)
async def _post_to_prompt_moderation_endpoint(self, payload: Mapping[str, object]) -> Mapping[str, Any]:
async def _post_to_prompt_moderation_endpoint(self, payload: Mapping[str, object]) -> _ModerationResponse:
"""Post a bare OpenAI request to the before_prompt webhook.
Returns ``{}`` (passthrough) or a synthetic chat.completion (block).
@ -1064,14 +1159,14 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
return await self._post_json(self.prompt_moderation_endpoint, payload, "Prompt moderation service")
@staticmethod
def _extract_prompt_refusal(service_response: Mapping[str, Any]) -> str | None:
def _extract_prompt_refusal(service_response: _ModerationResponse) -> str | None:
"""Return the refusal text when the prompt was blocked, else None.
The before_prompt webhook returns ``{}`` (passthrough) or a synthetic
chat.completion whose ``choices[0].message.content`` is the refusal
explanation.
"""
choices: Final[Sequence[_ServiceChoice] | None] = service_response.get("choices")
choices: Final = service_response.get("choices")
if not choices:
return None
message: Final = choices[0].get("message") or _EMPTY_MAPPING
@ -1080,7 +1175,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
@staticmethod
def _extract_response_block(
service_response: Mapping[str, Any],
service_response: _ModerationResponse,
all_tool_calls: Sequence[ChatCompletionMessageToolCall],
sent_content: str,
) -> BlockedResponseResult | None:
@ -1103,7 +1198,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
Expects service_response in OpenAI chat completion format:
{"choices": [{"message": {"tool_calls": [...], "content": "..."}}]}
"""
choices: Final[Sequence[_ServiceChoice]] = service_response.get("choices") or ()
choices: Final = service_response.get("choices") or ()
if not choices:
raise _MalformedToolBlockingResponseError("Response moderation service returned empty response")

148
litellm/integrations/websearch_interception/handler.py
View file

@ -10,7 +10,7 @@ import asyncio
import math
import uuid
from collections.abc import AsyncIterator, Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final, TypedDict, TypeVar, cast
from typing import TYPE_CHECKING, Any, Final, Literal, Never, TypedDict, TypeVar, cast
from typing_extensions import ReadOnly
@ -46,7 +46,13 @@ from litellm.types.integrations.websearch_interception import (
AnthropicServerToolUseBlock,
WebSearchInterceptionConfig,
)
from litellm.types.llms.openai import AllMessageValues
from litellm.types.llms.anthropic import AnthropicThinkingParam
from litellm.types.llms.openai import (
AllMessageValues,
ChatCompletionAudioParam,
ChatCompletionPredictionContentParam,
OpenAIWebSearchOptions,
)
from litellm.types.utils import (
AgenticLoopParams,
CallTypes,
@ -56,6 +62,8 @@ from litellm.types.utils import (
from litellm.utils import ProviderConfigManager
if TYPE_CHECKING:
from aiohttp import ClientSession
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.llms.base_llm.anthropic_messages.transformation import (
BaseAnthropicMessagesConfig,
@ -94,23 +102,98 @@ class _WebSearchSettingsView(TypedDict):
websearch_interception_params: WebSearchInterceptionConfig
class _SearchToolLitellmParams(TypedDict, total=False):
search_provider: ReadOnly[str | None]
class _SearchToolConfig(TypedDict, total=False):
search_tool_name: str
litellm_params: Mapping[str, object] | None
litellm_params: ReadOnly[_SearchToolLitellmParams | None]
class _DeploymentKwargsView(TypedDict):
"""Typed reads of the untyped request kwargs seen by the deployment hook."""
class _LitellmParamsProviderView(TypedDict, total=False):
custom_llm_provider: ReadOnly[str]
litellm_params: ReadOnly[Mapping[str, object]]
class _DeploymentCallKwargsView(TypedDict):
custom_llm_provider: ReadOnly[str]
litellm_params: ReadOnly[_LitellmParamsProviderView]
model: ReadOnly[str]
class _UserAuthView(TypedDict):
"""Typed read of the optional team attached to the caller's auth object."""
class _AcreateNamedParams(TypedDict, total=False):
metadata: ReadOnly[Never]
stop_sequences: ReadOnly[Never]
stream: ReadOnly[bool | None]
system: ReadOnly[str | None]
temperature: ReadOnly[float | None]
thinking: ReadOnly[Never]
tool_choice: ReadOnly[Never]
tools: ReadOnly[Never]
top_k: ReadOnly[int | None]
top_p: ReadOnly[float | None]
container: ReadOnly[Never]
team_id: ReadOnly[str | None]
class _AsearchNamedParams(TypedDict, total=False):
max_results: ReadOnly[int | None]
search_domain_filter: ReadOnly[Never]
max_tokens_per_page: ReadOnly[int | None]
country: ReadOnly[str | None]
api_key: ReadOnly[str | None]
api_base: ReadOnly[str | None]
timeout: ReadOnly[float | None]
extra_headers: ReadOnly[Never]
class _AcompletionNamedParams(TypedDict, total=False):
functions: ReadOnly[Never]
function_call: ReadOnly[str | None]
timeout: ReadOnly[float | None]
temperature: ReadOnly[float | None]
top_p: ReadOnly[float | None]
n: ReadOnly[int | None]
stream: ReadOnly[bool | None]
stream_options: ReadOnly[Never]
stop: ReadOnly[Never]
max_tokens: ReadOnly[int | None]
max_completion_tokens: ReadOnly[int | None]
modalities: ReadOnly[Never]
prediction: ReadOnly[ChatCompletionPredictionContentParam | None]
audio: ReadOnly[ChatCompletionAudioParam | None]
presence_penalty: ReadOnly[float | None]
frequency_penalty: ReadOnly[float | None]
logit_bias: ReadOnly[Never]
user: ReadOnly[str | None]
response_format: ReadOnly[Never]
seed: ReadOnly[int | None]
tools: ReadOnly[Never]
tool_choice: ReadOnly[Never]
parallel_tool_calls: ReadOnly[bool | None]
logprobs: ReadOnly[bool | None]
top_logprobs: ReadOnly[int | None]
deployment_id: ReadOnly[str | None]
reasoning_effort: ReadOnly[Literal["none", "minimal", "low", "medium", "high", "xhigh", "default"] | None]
verbosity: ReadOnly[Literal["low", "medium", "high"] | None]
safety_identifier: ReadOnly[str | None]
service_tier: ReadOnly[str | None]
store: ReadOnly[bool | None]
prompt_cache_key: ReadOnly[str | None]
base_url: ReadOnly[str | None]
api_version: ReadOnly[str | None]
api_key: ReadOnly[str | None]
model_list: ReadOnly[Never]
extra_headers: ReadOnly[Never]
thinking: ReadOnly[AnthropicThinkingParam | None]
web_search_options: ReadOnly[OpenAIWebSearchOptions | None]
include_server_side_tool_invocations: ReadOnly[bool | None]
shared_session: ReadOnly["ClientSession | None"]
enable_json_schema_validation: ReadOnly[bool | None]
_NO_ACREATE_NAMED: Final[_AcreateNamedParams] = {}
_NO_ASEARCH_NAMED: Final[_AsearchNamedParams] = {}
_NO_ACOMPLETION_NAMED: Final[_AcompletionNamedParams] = {}
class WebSearchInterceptionLogger(CustomLogger):
@ -312,17 +395,17 @@ class WebSearchInterceptionLogger(CustomLogger):
"""
# Check if this is for an enabled provider
# Try top-level kwargs first, then nested litellm_params, then derive from model name
kwargs_view: Final[_DeploymentKwargsView] = {
call_kwargs_view: Final[_DeploymentCallKwargsView] = {
"custom_llm_provider": kwargs.get("custom_llm_provider", ""),
"litellm_params": kwargs.get("litellm_params", {}),
"model": kwargs.get("model", ""),
}
custom_llm_provider = kwargs_view["custom_llm_provider"] or kwargs_view["litellm_params"].get(
custom_llm_provider = call_kwargs_view["custom_llm_provider"] or call_kwargs_view["litellm_params"].get(
"custom_llm_provider", ""
)
if not custom_llm_provider:
try:
_, custom_llm_provider, _, _ = litellm.get_llm_provider(model=kwargs_view["model"])
_, custom_llm_provider, _, _ = litellm.get_llm_provider(model=call_kwargs_view["model"])
except Exception:
custom_llm_provider = ""
if custom_llm_provider not in self.enabled_providers:
@ -1218,10 +1301,10 @@ class WebSearchInterceptionLogger(CustomLogger):
messages: list[dict],
tool_calls: list[dict],
thinking_blocks: list[dict],
anthropic_messages_optional_request_params: dict,
anthropic_messages_optional_request_params: Mapping[str, object],
logging_obj: "LiteLLMLoggingObj | None",
stream: bool,
kwargs: dict,
kwargs: Mapping[str, object],
) -> "AnthropicMessagesResponse | AsyncIterator[object]":
"""Legacy path: execute search + build patch + run follow-up call."""
request_patch, structured_results = await self._build_anthropic_request_patch(
@ -1229,9 +1312,9 @@ class WebSearchInterceptionLogger(CustomLogger):
messages=messages,
tool_calls=tool_calls,
thinking_blocks=thinking_blocks,
anthropic_messages_optional_request_params=anthropic_messages_optional_request_params,
anthropic_messages_optional_request_params=dict[str, object](anthropic_messages_optional_request_params),
logging_obj=logging_obj,
kwargs=kwargs,
kwargs=dict[str, object](kwargs),
)
if request_patch.messages is None:
raise ValueError("WebSearchInterception: missing follow-up messages")
@ -1246,12 +1329,14 @@ class WebSearchInterceptionLogger(CustomLogger):
if max_tokens is None:
max_tokens = cast(int, kwargs.get("max_tokens", 1024))
patch_kwargs: Final = dict[str, object](request_patch.kwargs)
response: AnthropicMessagesResponse | AsyncIterator[object] = await anthropic_messages.acreate(
max_tokens=max_tokens,
messages=request_patch.messages,
model=request_patch.model or model,
**_NO_ACREATE_NAMED,
**optional_params,
**request_patch.kwargs,
**patch_kwargs,
)
# Legacy path: the new path goes through the typed plan + core
@ -1393,12 +1478,13 @@ class WebSearchInterceptionLogger(CustomLogger):
search_tool: Final = self._select_search_tool_from_router(llm_router=llm_router)
search_provider: str | None = None
search_litellm_params: dict[str, Any] = {}
search_litellm_params: Mapping[str, object] = {}
search_tool_name: Final = self._selected_search_tool_name(search_tool=search_tool)
if search_tool is not None:
await self._authorize_search_tool(search_tool=search_tool, kwargs=kwargs)
search_litellm_params = dict(search_tool.get("litellm_params", {}) or {})
search_provider = search_litellm_params.get("search_provider")
tool_params: Final[_SearchToolLitellmParams] = search_tool.get("litellm_params", {}) or {}
search_litellm_params = dict[str, object](tool_params)
search_provider = tool_params.get("search_provider")
# Fallback to perplexity if no router or no search tools configured
if not search_provider:
@ -1426,12 +1512,15 @@ class WebSearchInterceptionLogger(CustomLogger):
if key != "search_provider" and value is not None
}
result: Final = (
await litellm.asearch(query=query, search_provider=search_provider, **search_kwargs)
await litellm.asearch(
query=query, search_provider=search_provider, **_NO_ASEARCH_NAMED, **search_kwargs
)
if search_metadata is None
else await litellm.asearch(
query=query,
search_provider=search_provider,
litellm_metadata=search_metadata,
**_NO_ASEARCH_NAMED,
**search_kwargs,
)
)
@ -1471,8 +1560,7 @@ class WebSearchInterceptionLogger(CustomLogger):
valid_token=user_api_key_auth,
)
auth_view: Final[_UserAuthView] = {"team_id": getattr(user_api_key_auth, "team_id", None)}
team_id: Final = auth_view["team_id"]
team_id: Final[str | None] = getattr(user_api_key_auth, "team_id", None)
if team_id:
from litellm.proxy.proxy_server import (
prisma_client,
@ -1587,10 +1675,10 @@ class WebSearchInterceptionLogger(CustomLogger):
model: str,
messages: list[dict],
tool_calls: list[dict],
optional_params: dict,
optional_params: Mapping[str, object],
logging_obj: "LiteLLMLoggingObj | None",
stream: bool,
kwargs: dict,
kwargs: Mapping[str, object],
response_format: str = "openai",
) -> "ModelResponse | CustomStreamWrapper":
"""Legacy path: execute search + build patch + run follow-up call."""
@ -1598,8 +1686,8 @@ class WebSearchInterceptionLogger(CustomLogger):
model=model,
messages=messages,
tool_calls=tool_calls,
optional_params=optional_params,
kwargs=kwargs,
optional_params=dict[str, object](optional_params),
kwargs=dict[str, object](kwargs),
response_format=response_format,
)
if request_patch.messages is None:
@ -1607,11 +1695,13 @@ class WebSearchInterceptionLogger(CustomLogger):
params: Final = dict(optional_params)
params.update(request_patch.optional_params)
params.pop("tool_choice", None)
patch_kwargs: Final = dict[str, object](request_patch.kwargs)
return await litellm.acompletion(
model=request_patch.model or model,
messages=request_patch.messages,
**_NO_ACOMPLETION_NAMED,
**params,
**request_patch.kwargs,
**patch_kwargs,
)
async def _build_chat_completion_request_patch(

28
litellm/litellm_core_utils/token_counter.py
View file

@ -693,6 +693,26 @@ def _count_document_tokens(
)
def _count_file_tokens(
file_value: object,
count_function: TokenCounterFunction,
use_default_image_token_count: bool,
) -> int:
"""An OpenAI `file` block is the chat-completions spelling of a document, so it prices like one."""
if not isinstance(file_value, Mapping):
return 0
filename: Final = file_value.get("filename")
file_data: Final = file_value.get("file_data")
name_tokens: Final = count_function(filename) if isinstance(filename, str) and filename else 0
if not isinstance(file_data, str) or not file_data:
return name_tokens
return name_tokens + calculate_img_tokens(
data=file_data,
mode="auto",
use_default_image_token_count=use_default_image_token_count,
)
def _count_anthropic_content(
content: Mapping[str, Any],
count_function: TokenCounterFunction,
@ -778,6 +798,12 @@ def _count_content_list(
use_default_image_token_count,
default_token_count,
)
elif c["type"] == "file":
num_tokens += _count_file_tokens(
c.get("file"),
count_function,
use_default_image_token_count,
)
elif c["type"] in ("tool_use", "tool_result"):
num_tokens += _count_anthropic_content(
c,
@ -807,7 +833,7 @@ def _count_content_list(
raise ValueError(
f"Invalid content item type: {content_type}. "
f"Expected str or dict with 'type' field "
f"(text, image_url, image, document, tool_use, tool_result, thinking, tool_reference)."
f"(text, image_url, image, document, file, tool_use, tool_result, thinking, tool_reference)."
)
return num_tokens
except Exception as e:

186
litellm/llms/anthropic/chat/guardrail_translation/handler.py
View file

@ -16,9 +16,9 @@ import json
from collections.abc import Mapping, Sequence
from copy import deepcopy
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Final, cast
from typing import TYPE_CHECKING, Any, Final, Protocol, cast, overload, runtime_checkable
from typing_extensions import assert_never
from typing_extensions import ReadOnly, TypedDict, assert_never
from litellm._logging import verbose_proxy_logger
from litellm.llms.anthropic.chat.transformation import AnthropicConfig
@ -58,6 +58,8 @@ from litellm.types.utils import (
)
if TYPE_CHECKING:
from fastapi import HTTPException
from litellm.integrations.custom_guardrail import (
CustomGuardrail,
ModifyResponseException,
@ -98,6 +100,48 @@ InputWriteBackTarget = (
)
class _SSEDelta(TypedDict, total=False):
type: ReadOnly[str]
text: ReadOnly[str]
stop_reason: ReadOnly[str | None]
class _SSEEventData(TypedDict, total=False):
delta: ReadOnly[_SSEDelta]
def _as_str_mapping(value: Mapping[str, object]) -> Mapping[str, object]:
return value
def _content_block_at(blocks: Sequence[object], index: int) -> object:
return blocks[index]
@runtime_checkable
class _ModelDumpBlock(Protocol):
def model_dump(self) -> Mapping[str, object]: ...
@runtime_checkable
class _TextAttrBlock(Protocol):
text: str
class _WritableMessage(Protocol):
@overload
def get(self, key: str, /) -> object | None: ...
@overload
def get(self, key: str, default: object, /) -> object: ...
def __setitem__(self, key: str, value: object, /) -> None: ...
def _as_writable(value: _WritableMessage) -> _WritableMessage:
return value
@dataclass(frozen=True, slots=True)
class ScannedText:
text: str
@ -126,7 +170,7 @@ class AnthropicMessagesHandler(BaseTranslation):
@staticmethod
def _build_streaming_usage_response(
responses_so_far: list[object],
responses_so_far: Sequence[object],
request_data: dict | None,
) -> ModelResponse | None:
chunks: Final = tuple(response for response in responses_so_far if isinstance(response, (str, bytes)))
@ -144,7 +188,7 @@ class AnthropicMessagesHandler(BaseTranslation):
self,
exc: "ModifyResponseException",
stream_started: bool = False,
responses_so_far: list[object] | None = None,
responses_so_far: Sequence[object] | None = None,
) -> list[bytes]:
"""
Build an Anthropic SSE sequence delivering the guardrail block message
@ -162,9 +206,22 @@ class AnthropicMessagesHandler(BaseTranslation):
would make Anthropic clients reject the stream.
"""
if stream_started:
return self._block_continuation_chunks(exc, responses_so_far or [])
return list(self._block_continuation_chunks(exc, responses_so_far or []))
return self._standalone_block_chunks(exc)
def build_stream_error_items(
self,
exc: "HTTPException",
responses_so_far: Sequence[Any] | None = None,
) -> Sequence[Any] | None:
from litellm.proxy.common_request_processing import (
serialize_http_exception_detail,
)
from litellm.proxy.guardrails.anthropic_sse import anthropic_sse_error_frames
message, _ = serialize_http_exception_detail(exc.detail)
return tuple(anthropic_sse_error_frames(message))
def _standalone_block_chunks(self, exc: "ModifyResponseException") -> list[bytes]:
import uuid
@ -187,7 +244,9 @@ class AnthropicMessagesHandler(BaseTranslation):
)
return list(FakeAnthropicMessagesStreamIterator(response=block_response))
def _block_continuation_chunks(self, exc: "ModifyResponseException", responses_so_far: list[object]) -> list[bytes]:
def _block_continuation_chunks(
self, exc: "ModifyResponseException", responses_so_far: Sequence[object]
) -> Sequence[bytes]:
"""Continue an already-started message: close the open content block,
append the block message as a new text block, then end the message --
without a second message_start."""
@ -199,7 +258,7 @@ class AnthropicMessagesHandler(BaseTranslation):
def _sse(event_type: str, payload: dict) -> bytes:
return f"event: {event_type}\ndata: {json.dumps(payload)}\n\n".encode()
output_tokens: Final = blocked_response_usage(getattr(exc, "original_response", None))["output_tokens"]
output_tokens: Final = blocked_response_usage(getattr(exc, "original_response", None)).get("output_tokens", 0)
open_index, max_index = self._content_block_state(responses_so_far)
new_index: Final = (max_index + 1) if max_index is not None else 0
chunks: list[bytes] = []
@ -237,7 +296,7 @@ class AnthropicMessagesHandler(BaseTranslation):
@staticmethod
def _content_block_state(
responses_so_far: list[object],
responses_so_far: Sequence[object],
) -> tuple[int | None, int | None]:
"""From the SSE chunks already sent to the client, return (open
content-block index or None, highest content-block index seen or None).
@ -263,7 +322,20 @@ class AnthropicMessagesHandler(BaseTranslation):
return open_index, max_index
@staticmethod
def _iter_sse_events(item: object) -> list[dict[str, object]]:
def _parse_sse_data_line(raw_line: str) -> tuple[Mapping[str, object], ...]:
line: Final = raw_line.strip()
if not line.startswith("data:"):
return ()
try:
parsed: Final[object] = json.loads(line[len("data:") :].strip())
except json.JSONDecodeError:
return ()
if not isinstance(parsed, dict):
return ()
return (_as_str_mapping(parsed),)
@staticmethod
def _iter_sse_events(item: object) -> Sequence[Mapping[str, object]]:
"""Yield the event-data dicts in one stream chunk.
Handles both formats this stream can carry (see
@ -271,24 +343,15 @@ class AnthropicMessagesHandler(BaseTranslation):
several events separated by a blank line -- and an already-parsed event
``dict``."""
if isinstance(item, dict):
return [item]
return (_as_str_mapping(item),)
if not isinstance(item, (bytes, bytearray)):
return []
events: Final[list[dict[str, object]]] = []
for block in item.decode("utf-8", errors="replace").split("\n\n"):
for line in block.split("\n"):
line = line.strip()
if not line.startswith("data:"):
continue
try:
parsed: str | int | float | bool | None | Sequence[object] | Mapping[str, object] = json.loads(
line[len("data:") :].strip()
)
except json.JSONDecodeError:
continue
if isinstance(parsed, dict):
events.append(parsed)
return events
return ()
return tuple(
event
for block in item.decode("utf-8", errors="replace").split("\n\n")
for line in block.split("\n")
for event in AnthropicMessagesHandler._parse_sse_data_line(line)
)
def _translate_to_openai(self, data: dict) -> ChatCompletionRequest:
"""Translate Anthropic request to OpenAI chat completion format."""
@ -321,7 +384,7 @@ class AnthropicMessagesHandler(BaseTranslation):
data: dict,
guardrail_to_apply: "CustomGuardrail",
litellm_logging_obj: "LiteLLMLoggingObj | None" = None,
) -> Any:
) -> Mapping[str, object]:
"""
Process input messages by applying guardrails to text content.
"""
@ -481,7 +544,7 @@ class AnthropicMessagesHandler(BaseTranslation):
@staticmethod
def _openai_system_message_to_anthropic(
message: dict[str, object],
message: Mapping[str, object],
) -> dict[str, object] | None: # mutable-ok: API message payload
"""Convert an OpenAI system message to the client's Anthropic-shaped entry."""
content: Final = message.get("content")
@ -561,7 +624,7 @@ class AnthropicMessagesHandler(BaseTranslation):
@staticmethod
def _defer_systems_inside_tool_exchanges(
structured_messages: list, # mutable-ok: API message payload
structured_messages: Sequence[Mapping[str, object]],
) -> list:
"""Hold a system row until the tool exchange around it completes so the call/result pair converts together."""
from litellm.litellm_core_utils.prompt_templates.factory import group_tool_exchanges
@ -755,7 +818,7 @@ class AnthropicMessagesHandler(BaseTranslation):
if scan_only_tool_results:
return EMPTY_EXTRACTED_INPUT
text_str: Final = content_item.get("text", None)
text_str: Final[str | None] = content_item.get("text")
return ExtractedInput(
scanned=(
() if text_str is None else (ScannedText(text_str, ContentBlockTextTarget(msg_idx, content_idx)),)
@ -805,7 +868,7 @@ class AnthropicMessagesHandler(BaseTranslation):
async def _apply_guardrail_responses_to_input(
self,
messages: list[dict[str, object]],
messages: Sequence[_WritableMessage],
responses: list[str],
scanned: tuple[ScannedText, ...],
) -> None:
@ -931,7 +994,7 @@ class AnthropicMessagesHandler(BaseTranslation):
litellm_logging_obj: "LiteLLMLoggingObj | None" = None,
user_api_key_dict: "UserAPIKeyAuth | None" = None,
request_data: dict | None = None,
) -> list[Any]:
) -> Sequence[object]:
"""
Process output streaming response by applying guardrails to text content.
@ -1027,7 +1090,7 @@ class AnthropicMessagesHandler(BaseTranslation):
return request_data
@staticmethod
def _get_response_content(response: object) -> list[Any]:
def _get_response_content(response: object) -> Sequence[object]:
"""Extract content list from a dict or object response."""
if isinstance(response, dict):
return response.get("content", []) or []
@ -1037,7 +1100,7 @@ class AnthropicMessagesHandler(BaseTranslation):
def _extract_from_content_blocks(
self,
response_content: list[Any],
response_content: Sequence[object],
texts_to_check: list[str],
images_to_check: list[str],
task_mappings: list[tuple[int, int | None]],
@ -1045,21 +1108,10 @@ class AnthropicMessagesHandler(BaseTranslation):
) -> None:
"""Extract text, images, and tool calls from content blocks."""
for content_idx, content_block in enumerate(response_content):
block_dict: dict[str, object] = {}
if isinstance(content_block, dict):
block_type = content_block.get("type")
block_dict = cast(dict[str, object], content_block)
elif hasattr(content_block, "type"):
block_type = getattr(content_block, "type", None)
if hasattr(content_block, "model_dump"):
block_dict = content_block.model_dump()
else:
block_dict = {
"type": block_type,
"text": getattr(content_block, "text", None),
}
else:
fields = self._output_block_fields(content_block)
if fields is None:
continue
block_type, block_dict = fields
if block_type in ["text", "tool_use"]:
self._extract_output_text_and_images(
@ -1071,6 +1123,21 @@ class AnthropicMessagesHandler(BaseTranslation):
tool_calls_to_check=tool_calls_to_check,
)
@staticmethod
def _output_block_fields(content_block: object) -> "tuple[object, Mapping[str, object]] | None":
if isinstance(content_block, dict):
block_dict: Final = _as_str_mapping(content_block)
return block_dict.get("type"), block_dict
if not hasattr(content_block, "type"):
return None
block_type: Final = getattr(content_block, "type", None)
if isinstance(content_block, _ModelDumpBlock):
return block_type, content_block.model_dump()
return block_type, {
"type": block_type,
"text": getattr(content_block, "text", None),
}
@staticmethod
def _build_guardrail_inputs(
texts_to_check: list[str],
@ -1093,7 +1160,7 @@ class AnthropicMessagesHandler(BaseTranslation):
inputs["model"] = response_model
return inputs
def get_streaming_string_so_far(self, responses_so_far: list[Any]) -> str:
def get_streaming_string_so_far(self, responses_so_far: Sequence[object]) -> str:
"""
Parse streaming responses and extract accumulated text content.
@ -1164,7 +1231,7 @@ class AnthropicMessagesHandler(BaseTranslation):
# Only process content_block_delta events
if event_type == "content_block_delta" and data_line:
try:
data = json.loads(data_line)
data: _SSEEventData = json.loads(data_line)
delta = data.get("delta", {})
if delta.get("type") == "text_delta":
text += delta.get("text", "")
@ -1176,7 +1243,7 @@ class AnthropicMessagesHandler(BaseTranslation):
return text
def _check_streaming_has_ended(self, responses_so_far: list[Any]) -> bool:
def _check_streaming_has_ended(self, responses_so_far: Sequence[object]) -> bool:
"""
Check if streaming response has ended by looking for non-null stop_reason.
@ -1227,7 +1294,7 @@ class AnthropicMessagesHandler(BaseTranslation):
# Check for message_delta event with stop_reason
if event_type == "message_delta" and data_line:
try:
data = json.loads(data_line)
data: _SSEEventData = json.loads(data_line)
delta = data.get("delta", {})
stop_reason = delta.get("stop_reason")
if stop_reason is not None:
@ -1271,7 +1338,7 @@ class AnthropicMessagesHandler(BaseTranslation):
def _extract_output_text_and_images(
self,
content_block: dict[str, object],
content_block: Mapping[str, object],
content_idx: int,
texts_to_check: list[str],
images_to_check: list[str],
@ -1294,7 +1361,7 @@ class AnthropicMessagesHandler(BaseTranslation):
task_mappings.append((content_idx, None))
# Extract tool calls
elif content_type == "tool_use":
elif content_type == "tool_use" and isinstance(content_block, dict):
tool_call: Final = AnthropicConfig.convert_tool_use_to_openai_format(
anthropic_tool_content=content_block,
index=content_idx,
@ -1319,7 +1386,7 @@ class AnthropicMessagesHandler(BaseTranslation):
content_idx = cast(int, mapping[0])
# Handle both dict and object responses
response_content: list[Any] = []
response_content: Sequence[object] = []
if isinstance(response, dict):
response_content = response.get("content", []) or []
elif hasattr(response, "content"):
@ -1335,14 +1402,15 @@ class AnthropicMessagesHandler(BaseTranslation):
if content_idx >= len(response_content):
continue
content_block = response_content[content_idx]
content_block = _content_block_at(response_content, content_idx)
# Verify it's a text block and update the text field
# Handle both dict and Pydantic object content blocks
if isinstance(content_block, dict):
if content_block.get("type") == "text":
cast(dict[str, object], content_block)["text"] = guardrail_response
block = _as_writable(content_block)
if block.get("type") == "text":
block["text"] = guardrail_response
elif hasattr(content_block, "type") and getattr(content_block, "type", None) == "text":
# Update Pydantic object's text attribute
if hasattr(content_block, "text"):
if isinstance(content_block, _TextAttrBlock):
content_block.text = guardrail_response

243
litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
View file

@ -13,10 +13,10 @@ Mirrors Anthropic's native ``compact_20260112`` for non-Anthropic providers:
"""
import re
from collections.abc import Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final, Literal, NotRequired, Optional, TypedDict, Union, cast
from collections.abc import Awaitable, Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, Protocol, TypeVar, Union, cast
from typing_extensions import ReadOnly
from typing_extensions import NotRequired, ReadOnly, TypedDict, Unpack
import litellm
from litellm._logging import verbose_logger
@ -29,6 +29,7 @@ from litellm.types.llms.anthropic import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.hooks.parallel_request_limiter_v3 import RateLimitDescriptor, RateLimitResponse
from litellm.router import Router
from litellm.types.llms.anthropic import (
AllAnthropicPassThroughMessageValues,
@ -84,6 +85,77 @@ _PROPAGATED_METADATA_KEYS: Final = (
_SUMMARY_TAG_RE: Final = re.compile(r"<summary>(.*?)</summary>", re.IGNORECASE | re.DOTALL)
_MsgT: Final = TypeVar("_MsgT", bound=Mapping[str, object])
def _as_object(value: object) -> object:
return value
def _is_tool_result_block(block: object) -> bool:
return isinstance(block, dict) and block.get("type") in ("tool_result",)
class _SummaryCallKwargs(TypedDict):
model: ReadOnly[str]
max_tokens: ReadOnly[int]
timeout: ReadOnly[float]
litellm_metadata: ReadOnly[Mapping[str, object]]
user: ReadOnly[NotRequired[str]]
allowed_model_region: ReadOnly[NotRequired[str]]
class _SummaryOptionalKwargs(TypedDict, total=False):
user: ReadOnly[str]
allowed_model_region: ReadOnly[str]
class _SummaryAcompletion(Protocol):
def __call__(
self,
*,
messages: Sequence[Mapping[str, object]],
**kwargs: Unpack[_SummaryCallKwargs], # kwargs-ok: forwarded verbatim to acompletion, which owns them
) -> "Awaitable[ModelResponse | CustomStreamWrapper]": ...
class _CreateRateLimitDescriptors(Protocol):
def __call__(
self,
*,
user_api_key_dict: "UserAPIKeyAuth",
data: Mapping[str, str],
rpm_limit_type: object,
tpm_limit_type: object,
model_has_failures: bool,
) -> "Sequence[RateLimitDescriptor]": ...
class _AddModelRateLimitDescriptor(Protocol):
def __call__(
self,
*,
user_api_key_dict: "UserAPIKeyAuth",
requested_model: str,
descriptors: "Sequence[RateLimitDescriptor]",
) -> None: ...
class _CreateOrgRateLimitDescriptors(Protocol):
def __call__(
self, user_api_key_dict: "UserAPIKeyAuth", requested_model: str | None = None
) -> "Sequence[RateLimitDescriptor]": ...
class _ShouldRateLimit(Protocol):
def __call__(
self,
*,
descriptors: "Sequence[RateLimitDescriptor]",
parent_otel_span: object,
read_only: bool,
) -> "Awaitable[RateLimitResponse]": ...
def _read_summary_model_setting() -> str | None:
"""Look up the configured summarization model from proxy general_settings."""
@ -159,11 +231,11 @@ async def _check_summary_model_access(
return True
key_models: Final = list(getattr(user_api_key_auth, "models", None) or [])
team_id: Final = getattr(user_api_key_auth, "team_id", None)
team_id: Final[str | None] = getattr(user_api_key_auth, "team_id", None)
team_model_aliases: Final = getattr(user_api_key_auth, "team_model_aliases", None)
team_models: Final = list(getattr(user_api_key_auth, "team_models", None) or [])
user_id: Final = getattr(user_api_key_auth, "user_id", None)
project_id: Final = getattr(user_api_key_auth, "project_id", None)
user_id: Final[str | None] = getattr(user_api_key_auth, "user_id", None)
project_id: Final[str | None] = getattr(user_api_key_auth, "project_id", None)
checks: Final[tuple[tuple[Literal["key", "team"], list[str]], ...]] = (
("key", key_models),
@ -372,7 +444,7 @@ async def _check_summary_model_budget(
return False
end_user_model_max_budget: Final = getattr(user_api_key_auth, "end_user_model_max_budget", None)
end_user_id: Final = getattr(user_api_key_auth, "end_user_id", None)
end_user_id: Final[str | None] = getattr(user_api_key_auth, "end_user_id", None)
if isinstance(end_user_model_max_budget, dict) and end_user_model_max_budget and end_user_id is not None:
try:
await model_max_budget_limiter.is_end_user_within_model_budget(
@ -424,40 +496,57 @@ async def _check_summary_model_rate_limit(
except Exception:
return True
limiter: Final = getattr(proxy_logging_obj, "max_parallel_request_limiter", None)
limiter: Final[object] = getattr(proxy_logging_obj, "max_parallel_request_limiter", None)
should_rate_limit_check: Final[_ShouldRateLimit | None] = getattr(limiter, "should_rate_limit", None)
create_descriptors: Final[_CreateRateLimitDescriptors | None] = getattr(
limiter, "_create_rate_limit_descriptors", None
)
add_team_descriptor: Final[_AddModelRateLimitDescriptor | None] = getattr(
limiter, "_add_team_model_rate_limit_descriptor_from_metadata", None
)
add_project_descriptor: Final[_AddModelRateLimitDescriptor | None] = getattr(
limiter, "_add_project_model_rate_limit_descriptor_from_metadata", None
)
create_org_descriptors: Final[_CreateOrgRateLimitDescriptors | None] = getattr(
limiter, "create_organization_rate_limit_descriptor", None
)
if (
limiter is None
or not hasattr(limiter, "should_rate_limit")
or not hasattr(limiter, "_create_rate_limit_descriptors")
or should_rate_limit_check is None
or create_descriptors is None
or add_team_descriptor is None
or add_project_descriptor is None
or create_org_descriptors is None
):
return True
try:
metadata: Final = getattr(user_api_key_auth, "metadata", None) or {}
metadata: Final[Mapping[str, object]] = getattr(user_api_key_auth, "metadata", None) or {}
data: Final = {"model": summary_model}
descriptors: Final = limiter._create_rate_limit_descriptors(
base_descriptors: Final = create_descriptors(
user_api_key_dict=user_api_key_auth,
data=data,
rpm_limit_type=metadata.get("rpm_limit_type"),
tpm_limit_type=metadata.get("tpm_limit_type"),
model_has_failures=False,
)
limiter._add_team_model_rate_limit_descriptor_from_metadata(
add_team_descriptor(
user_api_key_dict=user_api_key_auth,
requested_model=summary_model,
descriptors=descriptors,
descriptors=base_descriptors,
)
limiter._add_project_model_rate_limit_descriptor_from_metadata(
add_project_descriptor(
user_api_key_dict=user_api_key_auth,
requested_model=summary_model,
descriptors=descriptors,
descriptors=base_descriptors,
)
descriptors.extend(limiter.create_organization_rate_limit_descriptor(user_api_key_auth, summary_model))
descriptors: Final = (*base_descriptors, *create_org_descriptors(user_api_key_auth, summary_model))
if not descriptors:
return True
response: Final = await limiter.should_rate_limit(
parent_otel_span: Final[object] = getattr(user_api_key_auth, "parent_otel_span", None)
response: Final[RateLimitResponse] = await should_rate_limit_check(
descriptors=descriptors,
parent_otel_span=getattr(user_api_key_auth, "parent_otel_span", None),
parent_otel_span=parent_otel_span,
read_only=True,
)
except Exception as e:
@ -471,7 +560,7 @@ async def _check_summary_model_rate_limit(
def _find_latest_compaction_index(
messages: list[dict[str, object]],
messages: Sequence[Mapping[str, object]],
) -> tuple[int | None, int | None]:
"""Return (message_index, block_index) of the most recent compaction block.
@ -490,8 +579,8 @@ def _find_latest_compaction_index(
def _slice_around_compaction_block(
messages: list[dict[str, Any]],
) -> tuple[list[dict[str, object]], dict[str, object] | None]:
messages: Sequence[_MsgT],
) -> tuple[Sequence[_MsgT | dict[str, object]], dict[str, object] | None]:
"""Apply Anthropic's "drop everything before the compaction block" rule.
Returns ``(sliced_messages_with_compaction_block, compaction_block_dict)``
@ -506,19 +595,21 @@ def _slice_around_compaction_block(
original_msg: Final = messages[msg_idx]
original_content: Final = original_msg["content"]
compaction_block: Final = cast(dict[str, object], original_content[blk_idx])
if not isinstance(original_content, list):
return messages, None
original_blocks: Final = cast("Sequence[dict[str, object]]", original_content)
compaction_block: Final = original_blocks[blk_idx]
# Per Anthropic's contract everything before the compaction block is
# dropped, including earlier blocks within the same assistant message.
sliced_content: Final = list(original_content[blk_idx:])
sliced_content: Final = list(original_blocks[blk_idx:])
sliced_messages: Final[list[dict[str, object]]] = [{**original_msg, "content": sliced_content}]
sliced_messages.extend(messages[msg_idx + 1 :])
sliced_messages: Final = [{**original_msg, "content": sliced_content}, *messages[msg_idx + 1 :]]
return sliced_messages, compaction_block
def _strip_compaction_blocks(
messages: list[dict[str, object]],
messages: Sequence[dict[str, object]],
) -> list[dict[str, object]]:
"""Drop any ``compaction`` content blocks from messages.
@ -625,7 +716,7 @@ def _propagate_metadata(
def _count_effective_tokens(
model: str,
effective_messages: list[dict[str, object]],
effective_messages: Sequence[dict[str, object]],
compaction_block: CompactionBlock | None,
tools: list[dict[str, object]] | None,
system: str | list[dict[str, object]] | None = None,
@ -704,17 +795,18 @@ def _system_to_text(
return ""
if isinstance(system, str):
return system
parts: Final[list[str]] = []
for block in system:
if isinstance(block, dict) and block.get("type") == "text":
text = block.get("text")
if isinstance(text, str) and text:
parts.append(text)
return "\n".join(parts)
return "\n".join(
text
for block in system
if isinstance(block, dict)
and block.get("type") == "text"
and isinstance(text := block.get("text"), str)
and text
)
def _select_last_user_question(
messages: list[dict[str, object]],
messages: Sequence[dict[str, object]],
) -> list[dict[str, object]]:
"""Pick the most recent ``user`` turn that is a real question.
@ -729,16 +821,18 @@ def _select_last_user_question(
turns, or contained no user turns at all). The downstream call always
needs a non-empty user message.
"""
blocks: Sequence[object]
for msg in reversed(messages):
if msg.get("role") != "user":
continue
content = msg.get("content")
if isinstance(content, list):
filtered = [blk for blk in content if not (isinstance(blk, dict) and blk.get("type") == "tool_result")]
blocks = [*map(_as_object, content)]
filtered = [blk for blk in blocks if not _is_tool_result_block(blk)]
if not filtered:
# Purely tool_result — skip and look for an earlier turn.
continue
if len(filtered) < len(content):
if len(filtered) < len(blocks):
return [{**msg, "content": filtered}]
return [msg]
return [
@ -761,7 +855,7 @@ def _extract_summary_text(raw: str | None) -> str | None:
def _system_to_openai_message(
system: str | list[dict[str, Any]] | None,
) -> dict[str, object] | None:
) -> Mapping[str, object] | None:
"""Translate Anthropic-shaped ``system`` to an OpenAI system message.
Accepts a bare string or a list of Anthropic content blocks; returns
@ -772,17 +866,19 @@ def _system_to_openai_message(
if isinstance(system, str):
return {"role": "system", "content": system} if system else None
if isinstance(system, list):
parts = [block.get("text", "") for block in system if isinstance(block, dict) and block.get("type") == "text"]
parts: Final[tuple[str, ...]] = tuple(
block.get("text", "") for block in system if isinstance(block, dict) and block.get("type") == "text"
)
joined: Final = "\n\n".join(part for part in parts if part)
return {"role": "system", "content": joined} if joined else None
return None
def _build_summary_messages(
effective_messages: list[dict[str, object]],
effective_messages: Sequence[dict[str, object]],
prompt: str,
system: str | list[dict[str, object]] | None = None,
) -> list[dict[str, object]]:
) -> Sequence[Mapping[str, object]]:
"""Build the OpenAI-shape message list for the summary call.
The caller's ``system`` prompt is prepended (the default summarization
@ -810,7 +906,7 @@ def _build_summary_messages(
)
openai_messages = stripped
summary_messages: Final[list[dict[str, object]]] = []
summary_messages: Final[list[Mapping[str, object]]] = []
system_message: Final = _system_to_openai_message(system)
if system_message is not None:
summary_messages.append(system_message)
@ -845,35 +941,17 @@ def _append_text_to_content(content: object, extra_text: str) -> object:
if isinstance(content, str):
return f"{content}\n\n{extra_text}"
if isinstance(content, list):
appended: Final[list[object]] = [*content, {"type": "text", "text": extra_text}]
appended: Final[Sequence[object]] = [*map(_as_object, content), {"type": "text", "text": extra_text}]
return appended
return [content, {"type": "text", "text": extra_text}]
class _SummaryCallUserKwarg(TypedDict, total=False):
user: ReadOnly[object]
class _SummaryCallRegionKwarg(TypedDict, total=False):
allowed_model_region: ReadOnly[str]
class _SummaryCallKwargs(TypedDict):
model: ReadOnly[str]
messages: ReadOnly[list[dict[str, object]]]
max_tokens: ReadOnly[int]
timeout: ReadOnly[float]
litellm_metadata: ReadOnly[Mapping[str, object]]
user: NotRequired[ReadOnly[object]]
allowed_model_region: NotRequired[ReadOnly[str]]
async def _call_summary_model(
*,
summary_model: str,
summary_messages: list[dict[str, object]],
summary_messages: Sequence[Mapping[str, object]],
metadata: Mapping[str, object],
llm_router: Any,
llm_router: object,
allowed_model_region: str | None = None,
max_tokens: int = COMPACT_SUMMARY_MAX_TOKENS,
) -> Union["ModelResponse", "CustomStreamWrapper"]:
@ -909,28 +987,37 @@ async def _call_summary_model(
# than from ``litellm_metadata``, so without it the summary tokens would not
# debit the caller's end-user counters.
end_user_id: Final = metadata.get("user_api_key_end_user_id")
user_kwargs: Final = (
_SummaryOptionalKwargs(user=end_user_id)
if isinstance(end_user_id, str) and end_user_id
else _SummaryOptionalKwargs()
)
region_kwargs: Final = (
_SummaryOptionalKwargs(allowed_model_region=allowed_model_region)
if allowed_model_region is not None
else _SummaryOptionalKwargs()
)
call_kwargs: Final[_SummaryCallKwargs] = {
"model": summary_model,
"messages": summary_messages,
"max_tokens": max_tokens,
"timeout": COMPACT_SUMMARY_TIMEOUT_SECONDS,
"litellm_metadata": metadata,
**(_SummaryCallUserKwarg(user=end_user_id) if end_user_id else _SummaryCallUserKwarg()),
**(
_SummaryCallRegionKwarg(allowed_model_region=allowed_model_region)
if allowed_model_region is not None
else _SummaryCallRegionKwarg()
),
**user_kwargs,
**region_kwargs,
}
if llm_router is not None and hasattr(llm_router, "acompletion"):
return await llm_router.acompletion(**call_kwargs)
return await litellm.acompletion(**call_kwargs)
router_acompletion: Final[_SummaryAcompletion | None] = getattr(llm_router, "acompletion", None)
if llm_router is not None and router_acompletion is not None:
return await router_acompletion(messages=summary_messages, **call_kwargs)
return await litellm.acompletion(messages=[*summary_messages], **call_kwargs)
def _extract_response_text(response: Any) -> str | None:
def _extract_response_text(response: object) -> str | None:
try:
choice: Final = response.choices[0]
message: Final = choice.message
choices: Final[Sequence[object] | None] = getattr(response, "choices", None)
if choices is None:
return None
choice: Final = choices[0]
message: Final = getattr(choice, "message", None)
content: Final = getattr(message, "content", None)
if isinstance(content, str):
return content
@ -946,7 +1033,7 @@ def _extract_response_text(response: Any) -> str | None:
def _extract_usage(response: object) -> tuple[int, int]:
usage: Final = getattr(response, "usage", None)
usage: Final[object] = getattr(response, "usage", None)
if usage is None:
return 0, 0
return (

48
litellm/llms/base_llm/guardrail_translation/base_translation.py
View file

@ -1,8 +1,11 @@
from abc import ABC, abstractmethod
from collections.abc import Sequence
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, Final, Optional
if TYPE_CHECKING:
from fastapi import HTTPException
from litellm.integrations.custom_guardrail import (
CustomGuardrail,
ModifyResponseException,
@ -73,6 +76,31 @@ class BaseTranslation(ABC):
return transformed
@staticmethod
def merge_user_api_key_metadata_into_request(
request_data: dict[str, Any], # mutable-ok: proxy hooks share and mutate the request payload dict in place
user_api_key_dict: Optional["UserAPIKeyAuth"],
) -> None:
"""
Add the prefixed ``user_api_key_*`` metadata to the request's resolved
metadata bucket without overwriting existing keys.
Writes must go through ``get_or_create_metadata_bucket``: creating a
``litellm_metadata`` key on a route whose bucket is ``metadata`` (chat
completions) flips the bucket for every later metadata write, and spend
logging never sees those writes (e.g. guardrail_information).
"""
from litellm.litellm_core_utils.core_helpers import (
get_or_create_metadata_bucket,
)
user_metadata: Final = BaseTranslation.transform_user_api_key_dict_to_metadata(user_api_key_dict)
if not user_metadata:
return
_, metadata_bucket = get_or_create_metadata_bucket(request_data)
for key, value in user_metadata.items():
metadata_bucket.setdefault(key, value)
@abstractmethod
async def process_input_messages(
self,
@ -147,6 +175,26 @@ class BaseTranslation(ABC):
"""
return None
def build_stream_error_items(
self,
exc: "HTTPException",
responses_so_far: Sequence[Any] | None = None,
) -> Sequence[Any] | None:
"""
Build the stream items that surface a guardrail HTTPException (a block
with the default exception-on-block config, or a failed scan) after the
response has already started streaming, in this endpoint's wire format.
Called only once chunks have been sent: the HTTP status is gone, so the
failure must travel as an in-stream error frame. ``responses_so_far``
holds the chunks the client has already received, for formats whose
error frame continues the stream (e.g. sequence numbers).
Returns None when the format has no in-stream error frame; the caller
then re-raises ``exc``. Override in endpoint subclasses.
"""
return None
def get_structured_messages(self, data: dict) -> list["AllMessageValues"] | None:
"""
Convert request data to OpenAI-spec structured messages.

9
litellm/llms/bedrock/chat/converse_transformation.py
View file

@ -1631,6 +1631,11 @@ class AmazonConverseConfig(BaseConfig):
bedrock_tool_config["toolChoice"] = tool_choice_values
self._drop_tool_choice_type_conflicting_with_tool_config(additional_request_params)
config_block_entries: Final = tuple(
(config_name, config_class, inference_params.pop(config_name, None))
for config_name, config_class in self.get_config_blocks().items()
)
data: Final[CommonRequestObject] = {
"inferenceConfig": self._transform_inference_params(inference_params=inference_params),
}
@ -1641,9 +1646,7 @@ class AmazonConverseConfig(BaseConfig):
if system_content_blocks:
data["system"] = system_content_blocks
# Handle all config blocks
for config_name, config_class in self.get_config_blocks().items():
config_value = inference_params.pop(config_name, None)
for config_name, config_class, config_value in config_block_entries:
if config_value is not None:
data[config_name] = config_class(**config_value)

32
litellm/llms/openai/chat/guardrail_translation/handler.py
View file

@ -14,6 +14,7 @@ Pattern Overview:
This pattern can be replicated for other message formats (e.g., Anthropic).
"""
from collections.abc import Sequence
from typing import TYPE_CHECKING, Any, Final, Union, cast
import litellm
@ -46,6 +47,8 @@ from litellm.types.utils import (
)
if TYPE_CHECKING:
from fastapi import HTTPException
from litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
@ -382,11 +385,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
if "response" not in request_data:
request_data["response"] = response
# Add user API key metadata with prefixed keys
if "litellm_metadata" not in request_data:
user_metadata: Final = self.transform_user_api_key_dict_to_metadata(user_api_key_dict)
if user_metadata:
request_data["litellm_metadata"] = user_metadata
self.merge_user_api_key_metadata_into_request(request_data, user_api_key_dict)
inputs: Final = GenericGuardrailAPIInputs(texts=texts_to_check)
if images_to_check:
@ -555,11 +554,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
if "responses" not in request_data:
request_data["responses"] = responses_so_far
# Add user API key metadata with prefixed keys
if "litellm_metadata" not in request_data:
user_metadata: Final = self.transform_user_api_key_dict_to_metadata(user_api_key_dict)
if user_metadata:
request_data["litellm_metadata"] = user_metadata
self.merge_user_api_key_metadata_into_request(request_data, user_api_key_dict)
inputs: Final = GenericGuardrailAPIInputs(texts=texts_to_check)
if images_to_check:
@ -591,6 +586,18 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
return responses_so_far
def build_stream_error_items(
self,
exc: "HTTPException",
responses_so_far: Sequence[Any] | None = None,
) -> Sequence[Any] | None:
import json
from litellm.proxy.common_request_processing import sse_error_payload
_, error_obj = sse_error_payload(exc)
return (f'data: {{"error": {json.dumps(error_obj)}}}\n\n'.encode(),)
@staticmethod
def _accumulate_string_content_by_choice_index(
responses_so_far: list["ModelResponseStream"],
@ -653,10 +660,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
request_data = {"responses": responses_so_far}
elif "responses" not in request_data:
request_data["responses"] = responses_so_far
if "litellm_metadata" not in request_data:
user_metadata: Final = self.transform_user_api_key_dict_to_metadata(user_api_key_dict)
if user_metadata:
request_data["litellm_metadata"] = user_metadata
self.merge_user_api_key_metadata_into_request(request_data, user_api_key_dict)
inputs: Final = GenericGuardrailAPIInputs(texts=texts_to_check)
if responses_so_far and getattr(responses_so_far[0], "model", None):

1
litellm/llms/openai/common_utils.py
View file

@ -268,6 +268,7 @@ class BaseOpenAILLM:
"max_retries",
"organization",
"api_base",
"workload_identity_config",
)
openai_client_fields: Final = (
BaseOpenAILLM.get_openai_client_initialization_param_fields(client_type=client_type)

61
litellm/llms/openai/openai.py
View file

@ -51,6 +51,7 @@ from .common_utils import (
drop_params_from_unprocessable_entity_error,
is_output_token_limit_error,
)
from .workload_identity import resolve_openai_workload_identity_config
openaiOSeriesConfig: Final = OpenAIOSeriesConfig()
openAIGPT5Config: Final = OpenAIGPT5Config()
@ -349,6 +350,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
client: OpenAI | AsyncOpenAI | None = None,
shared_session: Optional["ClientSession"] = None,
) -> OpenAI | AsyncOpenAI | None:
workload_identity_config: Final = resolve_openai_workload_identity_config(api_key=api_key, api_base=api_base)
client_initialization_params: Final[dict] = locals()
if client is None:
if not isinstance(max_retries, int):
@ -364,28 +366,49 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
if cached_client:
if isinstance(cached_client, OpenAI) or isinstance(cached_client, AsyncOpenAI):
return cached_client
http_client: Final[httpx.Client | httpx.AsyncClient | None] = (
OpenAIChatCompletion._get_async_http_client(shared_session=shared_session)
if is_async
else OpenAIChatCompletion._get_sync_http_client()
)
if is_async:
_new_client: OpenAI | AsyncOpenAI = AsyncOpenAI(
api_key=api_key,
base_url=api_base,
http_client=http_client,
timeout=timeout,
max_retries=max_retries,
organization=organization,
async_http_client: Final = OpenAIChatCompletion._get_async_http_client(shared_session=shared_session)
http_client: httpx.Client | httpx.AsyncClient | None = async_http_client
_new_client: OpenAI | AsyncOpenAI = (
AsyncOpenAI(
workload_identity=workload_identity_config.to_sdk_workload_identity(),
base_url=api_base,
http_client=async_http_client,
timeout=timeout,
max_retries=max_retries,
organization=organization,
)
if workload_identity_config is not None
else AsyncOpenAI(
api_key=api_key,
base_url=api_base,
http_client=async_http_client,
timeout=timeout,
max_retries=max_retries,
organization=organization,
)
)
else:
_new_client = OpenAI(
api_key=api_key,
base_url=api_base,
http_client=http_client,
timeout=timeout,
max_retries=max_retries,
organization=organization,
sync_http_client: Final = OpenAIChatCompletion._get_sync_http_client()
http_client = sync_http_client
_new_client = (
OpenAI(
workload_identity=workload_identity_config.to_sdk_workload_identity(),
base_url=api_base,
http_client=sync_http_client,
timeout=timeout,
max_retries=max_retries,
organization=organization,
)
if workload_identity_config is not None
else OpenAI(
api_key=api_key,
base_url=api_base,
http_client=sync_http_client,
timeout=timeout,
max_retries=max_retries,
organization=organization,
)
)
## SAVE CACHE KEY

106
litellm/llms/openai/responses/count_tokens/transformation.py
View file

@ -4,7 +4,100 @@ OpenAI Responses API token counting transformation logic.
This module handles the transformation of requests to OpenAI's /v1/responses/input_tokens endpoint.
"""
from typing import Any, Final
from collections.abc import Mapping, Sequence
from typing import Any, Final, Literal
from typing_extensions import ReadOnly, TypedDict
class ResponsesInputTextPart(TypedDict):
type: ReadOnly[Literal["input_text"]]
text: ReadOnly[str]
class ResponsesInputImagePart(TypedDict):
type: ReadOnly[Literal["input_image"]]
image_url: ReadOnly[str]
detail: ReadOnly[str]
class ResponsesInputFilePart(TypedDict):
type: ReadOnly[Literal["input_file"]]
filename: ReadOnly[str]
file_data: ReadOnly[str]
ResponsesInputPart = ResponsesInputTextPart | ResponsesInputImagePart | ResponsesInputFilePart
ResponsesContentRole = Literal["user", "assistant"]
def _chat_image_block_to_responses_part(image_url: object) -> ResponsesInputImagePart | None:
url: Final = image_url.get("url") if isinstance(image_url, Mapping) else image_url
if not isinstance(url, str) or not url:
return None
detail: Final = image_url.get("detail") if isinstance(image_url, Mapping) else None
part: Final[ResponsesInputImagePart] = {
"type": "input_image",
"image_url": url,
"detail": detail if isinstance(detail, str) and detail else "auto",
}
return part
def _chat_file_block_to_responses_part(file_value: object) -> ResponsesInputFilePart | None:
"""Only an inline file round trips: OpenAI rejects `file_data` without the `filename` beside it."""
if not isinstance(file_value, Mapping):
return None
filename: Final = file_value.get("filename")
file_data: Final = file_value.get("file_data")
if not isinstance(filename, str) or not filename or not isinstance(file_data, str) or not file_data:
return None
part: Final[ResponsesInputFilePart] = {
"type": "input_file",
"filename": filename,
"file_data": file_data,
}
return part
def _chat_block_to_responses_part(block: object, role: ResponsesContentRole) -> ResponsesInputPart | None:
if isinstance(block, str):
bare: Final[ResponsesInputTextPart] = {"type": "input_text", "text": block}
return bare
if not isinstance(block, Mapping):
return None
match block.get("type"):
case "text":
text_value: Final = block.get("text")
text: Final[ResponsesInputTextPart] = {
"type": "input_text",
"text": text_value if isinstance(text_value, str) else "",
}
return text
case "image_url" if role == "user":
return _chat_image_block_to_responses_part(block.get("image_url"))
case "file" if role == "user":
return _chat_file_block_to_responses_part(block.get("file"))
case _:
return None
def chat_content_blocks_to_responses_content(
content: Sequence[object],
role: ResponsesContentRole,
) -> str | tuple[ResponsesInputPart, ...]:
"""Text-only content collapses to a joined string, which every role accepts and counts identically.
Only a user turn may carry an image or file part: the Responses API rejects any part but
output_text and refusal inside an assistant turn.
"""
parts: Final = tuple(
part for part in (_chat_block_to_responses_part(block, role) for block in content) if part is not None
)
if any(part["type"] != "input_text" for part in parts):
return parts
return "\n".join(part["text"] for part in parts if part["type"] == "input_text")
class OpenAICountTokensConfig:
@ -120,18 +213,13 @@ class OpenAICountTokensConfig:
instructions_parts.append("\n".join(text_parts))
elif role == "user":
if isinstance(content, list):
# Extract text from content blocks for Responses API
text_parts = []
for block in content:
if isinstance(block, dict) and block.get("type") == "text":
text_parts.append(block.get("text", ""))
elif isinstance(block, str):
text_parts.append(block)
content = "\n".join(text_parts)
content = chat_content_blocks_to_responses_content(content, "user")
input_items.append({"role": "user", "content": content})
elif role == "assistant":
# Map tool_calls to Responses API function_call items
tool_calls = msg.get("tool_calls")
if isinstance(content, list):
content = chat_content_blocks_to_responses_content(content, "assistant")
if content:
input_items.append({"role": "assistant", "content": content})
if tool_calls:

35
litellm/llms/openai/responses/guardrail_translation/handler.py
View file

@ -48,6 +48,8 @@ from litellm.types.llms.openai import (
AllMessageValues,
ChatCompletionToolCallChunk,
ChatCompletionToolParam,
ErrorEvent,
ErrorEventError,
OpenAIMcpServerTool,
ResponsesAPIStreamEvents,
)
@ -59,6 +61,8 @@ from litellm.types.responses.main import (
from litellm.types.utils import GenericGuardrailAPIInputs
if TYPE_CHECKING:
from fastapi import HTTPException
from litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.proxy._types import UserAPIKeyAuth
@ -80,6 +84,14 @@ class ResponsesStreamChunk(TypedDict, total=False):
text: ReadOnly[str]
def _next_stream_sequence_number(responses_so_far: Sequence[Any] | None) -> int:
sequence_numbers: Final = (
item.get("sequence_number") if isinstance(item, dict) else getattr(item, "sequence_number", None)
for item in reversed(responses_so_far or ())
)
return next((n + 1 for n in sequence_numbers if isinstance(n, int)), 0)
class OpenAIResponsesHandler(BaseTranslation):
"""
Handler for processing OpenAI Responses API with guardrails.
@ -620,6 +632,29 @@ class OpenAIResponsesHandler(BaseTranslation):
}
return responses_so_far[-1].get("type") in terminal_types
def build_stream_error_items(
self,
exc: "HTTPException",
responses_so_far: Sequence[Any] | None = None,
) -> Sequence[Any] | None:
from litellm.proxy.common_request_processing import (
serialize_http_exception_detail,
)
message, _ = serialize_http_exception_detail(exc.detail)
return (
ErrorEvent(
type=ResponsesAPIStreamEvents.ERROR,
sequence_number=_next_stream_sequence_number(responses_so_far),
error=ErrorEventError(
type="guardrail_error",
code=str(exc.status_code),
message=message,
param=None,
),
),
)
def get_streaming_string_so_far(self, responses_so_far: Sequence[ResponsesStreamChunk]) -> str:
"""
Get the string so far from the responses so far.

9
litellm/llms/openai/responses/transformation.py
View file

@ -21,6 +21,7 @@ from litellm.types.router import GenericLiteLLMParams
from litellm.types.utils import LlmProviders
from ..common_utils import OpenAIError
from ..workload_identity import get_workload_identity_bearer_token, resolve_openai_workload_identity_config
OPENAI_RESPONSES_API_MIN_MAX_OUTPUT_TOKENS: Final = 16
@ -392,6 +393,14 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
litellm_params = litellm_params or GenericLiteLLMParams()
api_key = litellm_params.api_key or litellm.api_key or litellm.openai_key or get_secret_str("OPENAI_API_KEY")
headers.setdefault("Content-Type", "application/json")
workload_identity_config: Final = (
resolve_openai_workload_identity_config(api_key=api_key, api_base=litellm_params.api_base)
if self.custom_llm_provider is LlmProviders.OPENAI
else None
)
if workload_identity_config is not None:
headers["Authorization"] = f"Bearer {get_workload_identity_bearer_token(workload_identity_config)}"
return headers
headers["Authorization"] = f"Bearer {api_key}"
return headers

100
litellm/llms/openai/workload_identity.py Normal file
View file

@ -0,0 +1,100 @@
from __future__ import annotations
from dataclasses import dataclass
from functools import lru_cache
from typing import TYPE_CHECKING, Final
from urllib.parse import urlparse
import litellm
from litellm.secret_managers.main import get_secret_str, normalize_nonempty_secret_str
from .common_utils import OpenAIError
if TYPE_CHECKING:
from collections.abc import Callable
from openai.auth import SubjectTokenProvider, WorkloadIdentity, WorkloadIdentityAuth
OPENAI_WIF_CLIENT_ID: Final = "litellm"
_OPENAI_API_HOST: Final = "api.openai.com"
_SDK_UPGRADE_MESSAGE: Final = (
"OpenAI workload identity federation requires openai>=2.32.0. "
"Upgrade the installed openai package to use OPENAI_IDENTITY_PROVIDER_ID / "
"OPENAI_SERVICE_ACCOUNT_ID / OPENAI_IDENTITY_TOKEN_FILE."
)
@dataclass(frozen=True, slots=True)
class OpenAIWorkloadIdentityConfig:
identity_provider_id: str
service_account_id: str
token_file: str
def to_sdk_workload_identity(self) -> WorkloadIdentity:
k8s_token_provider: Final = _load_sdk_k8s_token_provider()
workload_identity: Final[WorkloadIdentity] = {
"client_id": OPENAI_WIF_CLIENT_ID,
"identity_provider_id": self.identity_provider_id,
"service_account_id": self.service_account_id,
"provider": k8s_token_provider(self.token_file),
}
return workload_identity
def resolve_openai_workload_identity_config(
api_key: str | None,
api_base: str | None,
) -> OpenAIWorkloadIdentityConfig | None:
static_api_key: Final = normalize_nonempty_secret_str(api_key) or normalize_nonempty_secret_str(
get_secret_str("OPENAI_API_KEY")
)
if static_api_key is not None:
return None
effective_api_base: Final = (
api_base or litellm.api_base or get_secret_str("OPENAI_BASE_URL") or get_secret_str("OPENAI_API_BASE")
)
if not _targets_openai_api(effective_api_base):
return None
identity_provider_id: Final = get_secret_str("OPENAI_IDENTITY_PROVIDER_ID")
service_account_id: Final = get_secret_str("OPENAI_SERVICE_ACCOUNT_ID")
token_file: Final = get_secret_str("OPENAI_IDENTITY_TOKEN_FILE")
if not identity_provider_id or not service_account_id or not token_file:
return None
return OpenAIWorkloadIdentityConfig(
identity_provider_id=identity_provider_id,
service_account_id=service_account_id,
token_file=token_file,
)
def get_workload_identity_bearer_token(config: OpenAIWorkloadIdentityConfig) -> str:
return _workload_identity_auth(config).get_token()
def _targets_openai_api(api_base: str | None) -> bool:
if api_base is None:
return True
parsed: Final = urlparse(api_base)
return parsed.scheme == "https" and parsed.hostname == _OPENAI_API_HOST
@lru_cache(maxsize=16)
def _workload_identity_auth(config: OpenAIWorkloadIdentityConfig) -> WorkloadIdentityAuth:
sdk_workload_identity_auth: Final = _load_sdk_workload_identity_auth()
return sdk_workload_identity_auth(workload_identity=config.to_sdk_workload_identity())
def _load_sdk_workload_identity_auth() -> type[WorkloadIdentityAuth]:
try:
from openai.auth import WorkloadIdentityAuth as sdk_workload_identity_auth
except ImportError as e:
raise OpenAIError(status_code=500, message=_SDK_UPGRADE_MESSAGE) from e
return sdk_workload_identity_auth
def _load_sdk_k8s_token_provider() -> Callable[[str], SubjectTokenProvider]:
try:
from openai.auth import k8s_service_account_token_provider
except ImportError as e:
raise OpenAIError(status_code=500, message=_SDK_UPGRADE_MESSAGE) from e
return k8s_service_account_token_provider

67
litellm/llms/runwayml/videos/transformation.py
View file

@ -117,6 +117,10 @@ class RunwayMLVideoConfig(BaseVideoConfig):
def __init__(self):
super().__init__()
@staticmethod
def _parse_task_response(raw_response: httpx.Response) -> _RunwayTaskResponse:
return raw_response.json()
def get_supported_openai_params(self, model: str) -> list:
"""
Get the list of supported OpenAI parameters for video generation.
@ -141,7 +145,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
video_create_optional_params: VideoCreateOptionalRequestParams,
model: str,
drop_params: bool,
) -> dict:
) -> dict[str, object]:
"""
Map OpenAI parameters to RunwayML format.
@ -151,37 +155,44 @@ class RunwayMLVideoConfig(BaseVideoConfig):
- size -> ratio (convert "WIDTHxHEIGHT" to "WIDTH:HEIGHT")
- seconds -> duration (convert to integer)
"""
mapped_params: Final[dict[str, object]] = {}
supported_openai_params: Final = self.get_supported_openai_params(model)
return {
**self._prompt_image_param(video_create_optional_params),
**self._ratio_param(video_create_optional_params),
**self._duration_param(video_create_optional_params),
# Pass through other parameters that aren't OpenAI-specific
**{key: value for key, value in video_create_optional_params.items() if key not in supported_openai_params},
}
@staticmethod
def _prompt_image_param(video_create_optional_params: VideoCreateOptionalRequestParams) -> Mapping[str, object]:
# Handle input_reference parameter - map to promptImage
# RunwayML supports URLs and data URIs directly
if "input_reference" in video_create_optional_params:
input_reference: Final = video_create_optional_params["input_reference"]
# RunwayML supports URLs and data URIs directly
mapped_params["promptImage"] = input_reference
return {"promptImage": video_create_optional_params["input_reference"]}
return {}
@staticmethod
def _ratio_param(video_create_optional_params: VideoCreateOptionalRequestParams) -> Mapping[str, str]:
# Handle size parameter - convert "1280x720" to "1280:720"
if "size" in video_create_optional_params:
size: Final = video_create_optional_params["size"]
if isinstance(size, str) and "x" in size:
mapped_params["ratio"] = size.replace("x", ":")
return {"ratio": size.replace("x", ":")}
return {}
@staticmethod
def _duration_param(video_create_optional_params: VideoCreateOptionalRequestParams) -> Mapping[str, int]:
# Handle seconds parameter - convert to integer
if "seconds" in video_create_optional_params:
seconds: Final = video_create_optional_params["seconds"]
if seconds is not None:
try:
mapped_params["duration"] = int(float(seconds)) if isinstance(seconds, str) else int(seconds)
return {"duration": int(float(seconds)) if isinstance(seconds, str) else int(seconds)}
except (ValueError, TypeError):
# If conversion fails, use default duration
pass
# Pass through other parameters that aren't OpenAI-specific
supported_openai_params: Final = self.get_supported_openai_params(model)
for key, value in video_create_optional_params.items():
if key not in supported_openai_params:
mapped_params[key] = value
return mapped_params
return {}
def validate_environment(
self,
@ -236,7 +247,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
model: str,
prompt: str,
api_base: str,
video_create_optional_request_params: dict,
video_create_optional_request_params: dict[str, object],
litellm_params: GenericLiteLLMParams,
headers: dict,
) -> tuple[dict, RequestFiles, str]:
@ -406,20 +417,18 @@ class RunwayMLVideoConfig(BaseVideoConfig):
# Get task status to retrieve video URL
url: Final = f"{api_base}/tasks/{encoded_video_id}"
params: Final[dict[str, str]] = {}
return url, dict[str, str]()
return url, params
def _extract_video_url_from_response(self, response_data: dict[str, Any]) -> str:
def _extract_video_url_from_response(self, response_data: _RunwayTaskResponse) -> str:
"""
Helper method to extract video URL from RunwayML response.
Shared between sync and async transforms.
"""
# Extract video URL from the output field
video_url = None
if "output" in response_data and response_data["output"]:
output: Final = response_data["output"]
video_url = output[0] if isinstance(output, list) else output
raw_output: Final = response_data.get("output")
if raw_output:
video_url = raw_output if isinstance(raw_output, str) else raw_output[0]
if not video_url:
# Check if the video generation failed or is still processing
@ -453,7 +462,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
"output":["https://dnznrvs05pmza.cloudfront.net/.../video.mp4?_jwt=..."]
}
"""
response_data: Final = raw_response.json()
response_data: Final[_RunwayTaskResponse] = self._parse_task_response(raw_response)
video_url: Final = self._extract_video_url_from_response(response_data)
# Download the video from the CloudFront URL synchronously
@ -482,7 +491,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
"output":["https://dnznrvs05pmza.cloudfront.net/.../video.mp4?_jwt=..."]
}
"""
response_data: Final = raw_response.json()
response_data: Final[_RunwayTaskResponse] = self._parse_task_response(raw_response)
video_url: Final = self._extract_video_url_from_response(response_data)
# Download the video from the CloudFront URL asynchronously
@ -564,9 +573,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
# Construct the URL for task cancellation
url: Final = f"{api_base}/tasks/{encoded_video_id}/cancel"
data: Final[dict[str, str]] = {}
return url, data
return url, dict[str, str]()
def transform_video_delete_response(
self,
@ -604,9 +611,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
url: Final = f"{api_base}/tasks/{encoded_video_id}"
# Empty dict for GET request (no body)
data: Final[dict[str, str]] = {}
return url, data
return url, dict[str, str]()
def transform_video_status_retrieve_response(
self,

20
litellm/llms/vertex_ai/vertex_llm_base.py
View file

@ -27,6 +27,15 @@ from .common_utils import (
get_vertex_base_url,
)
def _graft_default_vertex_path(api_base: str, default_url: str) -> str:
parsed_api_base: Final = urlparse(api_base)
default_segments: Final = urlparse(default_url).path.lstrip("/").split("/")
graft_segments: Final = default_segments[1:] if default_segments[0] in ("v1", "v1beta1") else default_segments
grafted_path: Final = parsed_api_base.path.rstrip("/") + "/" + "/".join(graft_segments)
return parsed_api_base._replace(path=grafted_path).geturl()
GOOGLE_IMPORT_ERROR_MESSAGE: Final = (
"Google Cloud SDK not found. Install it with: pip install 'litellm[google]' or pip install google-cloud-aiplatform"
)
@ -621,8 +630,9 @@ class VertexBase:
Handles custom api_base for:
1. Gemini (Google AI Studio) - constructs /models/{model}:{endpoint}
2. Vertex AI with standard proxies - constructs {api_base}:{endpoint};
if api_base has no path (bare host), grafts the default vertex URL path onto it
2. Vertex AI with standard proxies - grafts the default vertex URL path onto the
api_base when its path is empty or only an API version (/v1, /v1beta1);
otherwise constructs {api_base}:{endpoint}
3. Vertex AI with PSC endpoints - constructs full path structure
{api_base}/v1/projects/{project}/locations/{location}/endpoints/{model}:{endpoint}
(only when use_psc_endpoint_format=True)
@ -669,10 +679,14 @@ class VertexBase:
)
elif urlparse(api_base).path in ("", "/"):
url = api_base.rstrip("/") + urlparse(url).path
elif urlparse(api_base).path.rstrip("/") in ("/v1", "/v1beta1") and "/projects/" in urlparse(url).path:
url = _graft_default_vertex_path(api_base=api_base, default_url=url)
else:
url = f"{api_base}:{endpoint}"
if stream is True:
url = url + "?alt=sse"
parsed_stream_url: Final = urlparse(url)
stream_query: Final = f"{parsed_stream_url.query}&alt=sse" if parsed_stream_url.query else "alt=sse"
url = parsed_stream_url._replace(query=stream_query).geturl()
return auth_header, url
def _get_token_and_url(

295
litellm/model_prices_and_context_window_backup.json
View file

File diff suppressed because it is too large Load diff

19
litellm/proxy/_types.py
View file

@ -422,6 +422,9 @@ class LiteLLMRoutes(enum.Enum):
"/responses/{response_id}/cancel",
"/v1/responses/{response_id}/cancel",
"/openai/v1/responses/{response_id}/cancel",
"/responses/input_tokens",
"/v1/responses/input_tokens",
"/openai/v1/responses/input_tokens",
# vector stores
"/vector_stores",
"/v1/vector_stores",
@ -2541,7 +2544,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
)
alerting: list | None = Field(
None,
description="List of alerting integrations. Today, just slack - `alerting: ['slack']`",
description="List of alerting integrations - e.g. `alerting: ['slack', 'webhook', 'email']`. 'slack' posts Slack-format messages to any Slack-compatible webhook (Slack, Rocket.Chat, Mattermost); 'webhook' posts structured JSON budget alerts to WEBHOOK_URL",
)
alert_types: list[AlertType] | None = Field(
None,
@ -3549,6 +3552,19 @@ class AllCallbacks(LiteLLMPydanticObjectBase):
)
class SpendLogsRouterMetadata(TypedDict):
"""
Router provenance stamped on spend logs for deployments flagged with
model_info.internal_router_model, correlating the requested model group
with the provider deployment that served the call
"""
requested_model: ReadOnly[str | None]
selected_model: ReadOnly[str | None]
selected_provider: ReadOnly[str | None]
router_correlation_id: ReadOnly[str | None]
class SpendLogsMetadata(TypedDict):
"""
Specific metadata k,v pairs logged to spendlogs for easier cost tracking
@ -3591,6 +3607,7 @@ class SpendLogsMetadata(TypedDict):
compression_savings: CompressionSavingsMetadata | None
autorouter_savings: ReadOnly[float | None] # stamped by the logging payload; None = not auto-routed
litellm_gateway_injected_cache: ReadOnly[str | None]
router_metadata: ReadOnly[SpendLogsRouterMetadata | None] # None = deployment not flagged internal_router_model
class SpendLogsPayload(TypedDict):

6
litellm/proxy/client/cli/README.md
View file

@ -489,7 +489,7 @@ lite codex exec "summarize the repo"
Each command resolves your LiteLLM key (logging in via SSO when none is stored and you are at a terminal; otherwise it expects `LITELLM_PROXY_API_KEY` or `--api-key`), checks the key against the proxy so bad credentials fail immediately instead of deep inside the agent, exports the environment variables the agent reads, then replaces itself with the agent process.
The right variables are picked per agent. Claude Code gets `ANTHROPIC_BASE_URL` (the proxy root, so it appends `/v1/messages`) and `ANTHROPIC_AUTH_TOKEN`, with any stray `ANTHROPIC_API_KEY` cleared so the proxy token wins. Codex and OpenCode get `OPENAI_BASE_URL` (the proxy plus `/v1`) and `OPENAI_API_KEY`. Codex ignores `OPENAI_BASE_URL`, so it is additionally pointed at the proxy through a custom provider passed as `-c` config overrides (HTTP/SSE Responses transport, since the proxy does not speak the Responses WebSocket protocol).
The right variables are picked per agent. Claude Code gets `ANTHROPIC_BASE_URL` (the proxy root, so it appends `/v1/messages`) and `ANTHROPIC_AUTH_TOKEN`, with any stray `ANTHROPIC_API_KEY` cleared so the proxy token wins, and `ENABLE_TOOL_SEARCH=true` (unless you already set it) so Claude Code keeps tool search on even though the base URL is a proxy rather than a first-party Anthropic host. Codex and OpenCode get `OPENAI_BASE_URL` (the proxy plus `/v1`) and `OPENAI_API_KEY`. Codex ignores `OPENAI_BASE_URL`, so it is additionally pointed at the proxy through a custom provider passed as `-c` config overrides (HTTP/SSE Responses transport, since the proxy does not speak the Responses WebSocket protocol).
Options (these belong to the wrapper, so put them before the agent's own flags):
@ -505,7 +505,7 @@ The credential is short-lived by design (default 24h, configurable via `LITELLM_
### Route Every Claude Code Session Through the Proxy
`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL, `env.ENABLE_TOOL_SEARCH` to `true` when that key is missing, and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
Two things need to already be true: you've run `lite login` (or `lite login --pkce`, whose key the helper renews on its own), since the apiKeyHelper depends on that stored token, and the proxy is already reachable, since `lite up` does not start one for you.
@ -529,7 +529,7 @@ Cursor is not supported: it has no equivalent file-based config to hot-patch thi
lite --base-url https://your-proxy.example.com login --config-claude
```
It writes the same two settings `lite up` does, `env.ANTHROPIC_BASE_URL` and `apiKeyHelper`, but persistently: there is no backup, nothing to restore, and no foreground process to keep alive. Every other key in `~/.claude/settings.json` is preserved, the file is created if it does not exist, and it is written atomically with owner-only permissions. Plain `lite login` is unchanged; nothing happens to your Claude Code config unless you pass the flag.
It writes the same settings `lite up` does, `env.ANTHROPIC_BASE_URL`, `env.ENABLE_TOOL_SEARCH`, and `apiKeyHelper`, but persistently: there is no backup, nothing to restore, and no foreground process to keep alive. Every other key in `~/.claude/settings.json` is preserved, the file is created if it does not exist, and it is written atomically with owner-only permissions. Plain `lite login` is unchanged; nothing happens to your Claude Code config unless you pass the flag.
Because the credential is reached through `apiKeyHelper` rather than copied into the file, a later `lite login` refreshes it with no further action: Claude Code re-runs the helper on every request and picks up whatever token the most recent login stored. Nothing secret is written to `settings.json`.

9
litellm/proxy/client/cli/commands/agents.py
View file

@ -13,6 +13,8 @@ from .auth import context_secret_vault, get_stored_api_key, login
ANTHROPIC_BASE_URL_ENV: Final = "ANTHROPIC_BASE_URL"
ANTHROPIC_AUTH_TOKEN_ENV: Final = "ANTHROPIC_AUTH_TOKEN"
ANTHROPIC_API_KEY_ENV: Final = "ANTHROPIC_API_KEY"
ENABLE_TOOL_SEARCH_ENV: Final = "ENABLE_TOOL_SEARCH"
ENABLE_TOOL_SEARCH_VALUE: Final = "true"
OPENAI_BASE_URL_ENV: Final = "OPENAI_BASE_URL"
OPENAI_API_KEY_ENV: Final = "OPENAI_API_KEY"
@ -61,7 +63,10 @@ def build_agent_env(
Anthropic clients (Claude Code) append /v1/messages to ANTHROPIC_BASE_URL,
so it stays the bare proxy root; OpenAI clients (Codex, OpenCode) expect the
/v1 suffix on OPENAI_BASE_URL. ANTHROPIC_API_KEY is dropped so a stray
Anthropic key cannot win over the bearer token we set.
Anthropic key cannot win over the bearer token we set. ENABLE_TOOL_SEARCH
defaults to true because Claude Code turns tool search off when
ANTHROPIC_BASE_URL is not a first-party Anthropic host; a value already in
the environment is left alone.
"""
env: Final = dict(base_env)
root: Final = base_url.rstrip("/")
@ -69,6 +74,8 @@ def build_agent_env(
env[ANTHROPIC_BASE_URL_ENV] = root
env[ANTHROPIC_AUTH_TOKEN_ENV] = api_key
env.pop(ANTHROPIC_API_KEY_ENV, None)
if ENABLE_TOOL_SEARCH_ENV not in env:
env[ENABLE_TOOL_SEARCH_ENV] = ENABLE_TOOL_SEARCH_VALUE
if PROFILE_OPENAI in profiles:
env[OPENAI_BASE_URL_ENV] = root + "/v1"
env[OPENAI_API_KEY_ENV] = api_key

3
litellm/proxy/client/cli/commands/autoroute/settings.py
View file

@ -9,6 +9,8 @@ API_KEY_HELPER_KEY: Final = "apiKeyHelper"
ANTHROPIC_API_KEY_KEY: Final = "ANTHROPIC_API_KEY"
ANTHROPIC_AUTH_TOKEN_KEY: Final = "ANTHROPIC_AUTH_TOKEN"
ANTHROPIC_BASE_URL_KEY: Final = "ANTHROPIC_BASE_URL"
ENABLE_TOOL_SEARCH_KEY: Final = "ENABLE_TOOL_SEARCH"
ENABLE_TOOL_SEARCH_VALUE: Final = "true"
# Force every one of Claude Code's own model tiers to request the auto-router by name.
# Router's auto-router registry is keyed by the literal requested model string
# (litellm/router.py:10711-10717) with no wildcard/pattern resolution, so a bare "*"
@ -34,6 +36,7 @@ def merge_claude_settings_static_token(
raw_env: Final = settings.get(ENV_KEY, {})
base_env: Final = raw_env if isinstance(raw_env, dict) else {}
env: Final[dict[str, JsonValue]] = {
ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
**base_env,
ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
ANTHROPIC_AUTH_TOKEN_KEY: auth_token,

11
litellm/proxy/client/cli/commands/claude_settings.py
View file

@ -21,6 +21,8 @@ ENV_KEY: Final = "env"
API_KEY_HELPER_KEY: Final = "apiKeyHelper"
ANTHROPIC_BASE_URL_KEY: Final = "ANTHROPIC_BASE_URL"
ANTHROPIC_API_KEY_KEY: Final = "ANTHROPIC_API_KEY"
ENABLE_TOOL_SEARCH_KEY: Final = "ENABLE_TOOL_SEARCH"
ENABLE_TOOL_SEARCH_VALUE: Final = "true"
CLAUDE_SETTINGS_PATH: Final = Path.home() / ".claude" / "settings.json"
BACKUP_PATH: Final = Path.home() / ".litellm" / "claude_settings_backup.json"
@ -70,12 +72,15 @@ def merge_claude_settings(
Only env.ANTHROPIC_BASE_URL and the top-level apiKeyHelper are overridden; a
stray env.ANTHROPIC_API_KEY is dropped so it cannot outrank the helper-issued
token (same reasoning as build_agent_env in agents.py). Every other key is
preserved untouched.
token (same reasoning as build_agent_env in agents.py). ENABLE_TOOL_SEARCH
defaults to true because Claude Code turns tool search off when
ANTHROPIC_BASE_URL is not a first-party Anthropic host; an existing value is
left alone. Every other key is preserved untouched.
"""
raw_env: Final = settings.get(ENV_KEY, {})
base_env: Final = raw_env if isinstance(raw_env, dict) else {}
env: Final = {
ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
**{key: value for key, value in base_env.items() if key != ANTHROPIC_API_KEY_KEY},
ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
}
@ -144,6 +149,8 @@ __all__ = (
"AUTOROUTE_BACKUP_PATH",
"BACKUP_PATH",
"CLAUDE_SETTINGS_PATH",
"ENABLE_TOOL_SEARCH_KEY",
"ENABLE_TOOL_SEARCH_VALUE",
"ENV_KEY",
"SETTINGS_FILE_OWNERS",
"ClaudeSettingsError",

151
litellm/proxy/common_request_processing.py
View file

@ -502,7 +502,7 @@ def _as_success_dispatcher(logging_obj: _DispatchesSuccessHandlers) -> _Dispatch
return logging_obj
def _serialize_http_exception_detail(
def serialize_http_exception_detail(
detail: object,
) -> tuple[str, dict | None]:
"""
@ -535,7 +535,7 @@ def _serialize_http_exception_detail(
def proxy_exception_from_http_exception(exc: HTTPException, headers: dict[str, str]) -> ProxyException:
raw_detail: Final = _getattr_object(exc, "detail", str(exc))
message, structured_fields = _serialize_http_exception_detail(raw_detail)
message, structured_fields = serialize_http_exception_detail(raw_detail)
existing_fields: Final = getattr(exc, "provider_specific_fields", None) or {}
merged_fields: Final = {**existing_fields, **structured_fields} if structured_fields else (existing_fields or None)
return ProxyException(
@ -818,7 +818,7 @@ async def _buffer_first_chunk_honoring_disconnect(
raise _ClientDisconnectedBeforeFirstChunk()
def _sse_error_payload(exc: BaseException) -> tuple[int, Mapping[str, object]]:
def sse_error_payload(exc: BaseException) -> tuple[int, Mapping[str, object]]:
"""Build the ProxyException-shaped ``{"error": ...}`` body used in SSE error frames.
Matches ``ProxyException.to_dict()`` so streaming and non-streaming error frames
@ -827,7 +827,7 @@ def _sse_error_payload(exc: BaseException) -> tuple[int, Mapping[str, object]]:
# Preserve status code from HTTPException (e.g. guardrail blocks)
error_status: Final = getattr(exc, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR)
raw_detail: Final = _getattr_object(exc, "detail", "Error processing stream start")
message, structured_fields = _serialize_http_exception_detail(raw_detail)
message, structured_fields = serialize_http_exception_detail(raw_detail)
existing_fields: Final = getattr(exc, "provider_specific_fields", None) or {}
merged_fields: Final = {**existing_fields, **structured_fields} if structured_fields else (existing_fields or None)
@ -942,7 +942,7 @@ async def create_response(
# Unexpected error consuming first chunk.
verbose_proxy_logger.exception("Error consuming first chunk from generator: %s", e)
error_status, error_obj = _sse_error_payload(e)
error_status, error_obj = sse_error_payload(e)
async def error_gen_message() -> AsyncGenerator[str, None]:
for frame in _sse_error_frames(error_obj):
@ -1119,7 +1119,7 @@ async def open_sse_before_first_byte(
# would never fire and the failure would go unaudited. The hook
# also gets to sanitize what reaches the client, by returning or
# raising a replacement, so its answer decides the frame.
_, error_obj = _sse_error_payload(await _sanitized_late_failure(exc, on_late_failure))
_, error_obj = sse_error_payload(await _sanitized_late_failure(exc, on_late_failure))
for frame in _sse_error_frames(error_obj):
yield frame.encode()
return
@ -2380,53 +2380,14 @@ class ProxyBaseLLMRequestProcessing:
if requested_model_from_client:
self.data["_litellm_client_requested_model"] = requested_model_from_client
# Streaming: attach a closure that fires after all guardrail
# end-of-stream blocks complete. CSW.__anext__ stores the
# assembled response on logging_obj; the outer consumer
# (ProxyLogging._fire_deferred_stream_logging) fires the
# closure after the full streaming pipeline finishes.
# The closure runs non-apply_guardrail hooks on the
# assembled response, then fires success logging.
# Only for CustomStreamWrapper — raw async generators from
# passthrough routes bypass CSW and would orphan the closure.
from litellm.litellm_core_utils.streaming_handler import (
CustomStreamWrapper,
)
if _post_call_guardrails_active and isinstance(response, CustomStreamWrapper):
# Intentionally a live reference (not a copy) — mirrors
# ProxyLogging.post_call_success_hook which also mutates
# data["guardrail_to_apply"] during iteration.
_captured_data: Final = self.data
_captured_user_api_key_dict: Final = user_api_key_dict
_captured_logging_obj: Final = logging_obj
async def _on_deferred_stream_complete(assembled_response: object, cache_hit: object) -> None:
await ProxyBaseLLMRequestProcessing._run_deferred_stream_guardrails(
captured_data=_captured_data,
captured_user_api_key_dict=_captured_user_api_key_dict,
captured_logging_obj=_captured_logging_obj,
assembled_response=assembled_response,
cache_hit=cache_hit,
)
logging_obj._on_deferred_stream_complete = _on_deferred_stream_complete
elif (
_post_call_guardrails_active
and route_type == "anthropic_messages"
and self._is_streaming_response(response)
):
from litellm.litellm_core_utils.logging_worker import (
GLOBAL_LOGGING_WORKER,
if _post_call_guardrails_active:
self._arm_deferred_stream_dispatch(
response=response,
route_type=route_type,
user_api_key_dict=user_api_key_dict,
logging_obj=logging_obj,
)
async def _on_deferred_native_stream_complete(
logging_coroutine: Coroutine[object, object, object],
) -> None:
GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(async_coroutine=logging_coroutine)
logging_obj._on_deferred_stream_complete = _on_deferred_native_stream_complete
if route_type == "allm_passthrough_route":
# Check if response is an async generator
if self._is_streaming_response(response):
@ -3096,6 +3057,94 @@ class ProxyBaseLLMRequestProcessing:
except Exception as e:
verbose_proxy_logger.exception("Error firing deferred logging: %s", e)
def _arm_deferred_stream_dispatch(
self,
response: object,
route_type: str,
user_api_key_dict: "UserAPIKeyAuth",
logging_obj: LiteLLMLoggingObj,
) -> None:
"""
Streaming with post-call guardrails active: attach a closure that
ProxyLogging._fire_deferred_stream_logging fires after all guardrail
end-of-stream blocks complete, so the spend log sees
guardrail_information.
Three closure shapes, matching who owns logging for the stream:
- CustomStreamWrapper (chat completions) stores
(assembled_response, cache_hit); the closure also runs
non-apply_guardrail post-call hooks via
_run_deferred_stream_guardrails.
- Bridged /v1/responses (LiteLLMCompletionStreamingIterator) shares
its inner CustomStreamWrapper's logging_obj, so it stores the same
(assembled_response, cache_hit) shape; the closure only dispatches
success logging, matching the route's pre-existing hook surface.
- Native anthropic_messages/aresponses iterators store a single
ready-made logging coroutine to enqueue.
Raw async generators from passthrough routes bypass all three and
would orphan the closure, so they are not armed here.
The router wraps iterators that cannot carry _hidden_params in
HiddenParamsAsyncIteratorWrapper, so class sniffing runs on the
unwrapped inner iterator.
"""
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
from litellm.router_utils.add_retry_fallback_headers import HiddenParamsAsyncIteratorWrapper
unwrapped: Final = response._inner if isinstance(response, HiddenParamsAsyncIteratorWrapper) else response
if isinstance(unwrapped, CustomStreamWrapper):
# Intentionally a live reference (not a copy) — mirrors
# ProxyLogging.post_call_success_hook which also mutates
# data["guardrail_to_apply"] during iteration.
_captured_data: Final = self.data
_captured_user_api_key_dict: Final = user_api_key_dict
_captured_logging_obj: Final = logging_obj
async def _on_deferred_stream_complete(assembled_response: object, cache_hit: object) -> None:
await ProxyBaseLLMRequestProcessing._run_deferred_stream_guardrails(
captured_data=_captured_data,
captured_user_api_key_dict=_captured_user_api_key_dict,
captured_logging_obj=_captured_logging_obj,
assembled_response=assembled_response,
cache_hit=cache_hit,
)
logging_obj._on_deferred_stream_complete = _on_deferred_stream_complete
return
if route_type not in ("anthropic_messages", "aresponses") or not self._is_streaming_response(response):
return
from litellm.responses.litellm_completion_transformation.streaming_iterator import (
LiteLLMCompletionStreamingIterator,
)
if isinstance(unwrapped, LiteLLMCompletionStreamingIterator):
_captured_bridge_logging_obj: Final = logging_obj
async def _on_deferred_bridged_stream_complete(assembled_response: object, cache_hit: object) -> None:
await _as_success_dispatcher(_captured_bridge_logging_obj).dispatch_success_handlers(
assembled_response,
cache_hit=cache_hit,
start_time=None,
end_time=None,
prefer_async_handlers=True,
)
logging_obj._on_deferred_stream_complete = _on_deferred_bridged_stream_complete
return
from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
async def _on_deferred_native_stream_complete(
logging_coroutine: Coroutine[object, object, object],
) -> None:
GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(async_coroutine=logging_coroutine)
logging_obj._on_deferred_stream_complete = _on_deferred_native_stream_complete
@staticmethod
async def _run_deferred_stream_guardrails(
captured_data: dict,

101
litellm/proxy/db/tool_registry_writer.py
View file

@ -6,9 +6,11 @@ Admins use the management endpoints to read and update input_policy / output_pol
"""
import uuid
from collections.abc import Mapping
from collections.abc import Mapping, Sequence
from datetime import datetime, timezone
from typing import TYPE_CHECKING, Any, Final
from typing import TYPE_CHECKING, Final, Protocol
from pydantic import TypeAdapter
from litellm._logging import verbose_proxy_logger
from litellm.proxy._types import ToolDiscoveryQueueItem
@ -27,6 +29,13 @@ if TYPE_CHECKING:
from litellm.proxy.utils import PrismaClient
class _ModelDumpMethod(Protocol):
def __call__(self) -> Mapping: ...
_ROW_DICT: Final = TypeAdapter(dict)
def _tool_table_actions(prisma_client: "PrismaClient") -> "TableActions[prisma_db_models.LiteLLM_ToolTable]":
table: Final[TableActions[prisma_db_models.LiteLLM_ToolTable]] = ToolRepository(prisma_client).table
return table
@ -41,33 +50,35 @@ def _object_permission_table_actions(
return table
def _row_to_model(row: dict | Any) -> LiteLLM_ToolTableRow:
def _row_to_model(row: object) -> LiteLLM_ToolTableRow:
"""Convert a Prisma model instance or dict to LiteLLM_ToolTableRow."""
model_dump: Final = getattr(row, "model_dump", None)
model_dump: Final[_ModelDumpMethod | None] = getattr(row, "model_dump", None)
if callable(model_dump):
row = model_dump()
elif not isinstance(row, dict):
row = {
k: getattr(row, k, None)
for k in (
"tool_id",
"tool_name",
"origin",
"input_policy",
"output_policy",
"call_count",
"assignments",
"key_hash",
"team_id",
"key_alias",
"user_agent",
"last_used_at",
"created_at",
"updated_at",
"created_by",
"updated_by",
)
}
row = _ROW_DICT.validate_python(
{
k: getattr(row, k, None)
for k in (
"tool_id",
"tool_name",
"origin",
"input_policy",
"output_policy",
"call_count",
"assignments",
"key_hash",
"team_id",
"key_alias",
"user_agent",
"last_used_at",
"created_at",
"updated_at",
"created_by",
"updated_by",
)
}
)
return LiteLLM_ToolTableRow(
tool_id=row.get("tool_id", ""),
tool_name=row.get("tool_name", ""),
@ -190,7 +201,7 @@ async def update_tool_policy(
_updated_by: Final = updated_by or "system"
now: Final = datetime.now(timezone.utc)
create_data: Final[dict[str, object]] = {
create_data: Final[Mapping[str, str | datetime]] = {
"tool_id": str(uuid.uuid4()),
"tool_name": tool_name,
"input_policy": input_policy or "untrusted",
@ -200,14 +211,16 @@ async def update_tool_policy(
"created_at": now,
"updated_at": now,
}
update_data: Final[dict[str, object]] = {
"updated_by": _updated_by,
"updated_at": now,
update_data: Final[Mapping[str, str | datetime]] = {
key: value
for key, value in (
("updated_by", _updated_by),
("updated_at", now),
("input_policy", input_policy),
("output_policy", output_policy),
)
if value is not None
}
if input_policy is not None:
update_data["input_policy"] = input_policy
if output_policy is not None:
update_data["output_policy"] = output_policy
await _tool_table_actions(prisma_client).upsert(
where={"tool_name": tool_name},
@ -338,7 +351,7 @@ class ToolPolicyRegistry:
self._blocked_tools_by_op_id = {}
for row in perms:
op_id = getattr(row, "object_permission_id", None)
blocked = getattr(row, "blocked_tools", None) or []
blocked: Sequence[str] = getattr(row, "blocked_tools", None) or []
if op_id:
self._blocked_tools_by_op_id[op_id] = list(blocked)
@ -370,10 +383,12 @@ class ToolPolicyRegistry:
"""
if not tool_names:
return {}
blocked: Final[set[str]] = set()
for op_id in (object_permission_id, team_object_permission_id):
if op_id and op_id.strip():
blocked.update(self._blocked_tools_by_op_id.get(op_id.strip(), []))
blocked: Final[frozenset[str]] = frozenset(
tool
for op_id in (object_permission_id, team_object_permission_id)
if op_id and op_id.strip()
for tool in self._blocked_tools_by_op_id.get(op_id.strip(), [])
)
result: Final[dict[str, str]] = {}
for name in tool_names:
if name in blocked:
@ -408,13 +423,12 @@ async def add_tool_to_object_permission_blocked(
)
if row is None:
return False
current: Final = list(getattr(row, "blocked_tools", []) or [])
current: Final[Sequence[str]] = getattr(row, "blocked_tools", []) or []
if tool_name in current:
return True
current.append(tool_name)
await _object_permission_table_actions(prisma_client).update(
where={"object_permission_id": object_permission_id},
data={"blocked_tools": current},
data={"blocked_tools": [*current, tool_name]},
)
return True
except Exception as e:
@ -436,13 +450,12 @@ async def remove_tool_from_object_permission_blocked(
)
if row is None:
return False
current = list(getattr(row, "blocked_tools", []) or [])
current: Final[Sequence[str]] = getattr(row, "blocked_tools", []) or []
if tool_name not in current:
return False
current = [t for t in current if t != tool_name]
await _object_permission_table_actions(prisma_client).update(
where={"object_permission_id": object_permission_id},
data={"blocked_tools": current},
data={"blocked_tools": [t for t in current if t != tool_name]},
)
return True
except Exception as e:

53
litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py
View file

@ -18,6 +18,7 @@ import time
from collections.abc import AsyncGenerator, Mapping, Sequence
from datetime import datetime, timezone
from itertools import accumulate, groupby
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, ClassVar, Final, Literal, NamedTuple, Optional, cast
import httpx
@ -42,7 +43,7 @@ from litellm.llms.custom_httpx.http_handler import (
httpxSpecialProvider,
)
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.common_request_processing import _serialize_http_exception_detail
from litellm.proxy.common_request_processing import serialize_http_exception_detail
from litellm.proxy.common_utils.sse_keepalive import keepalive_ping_has_fired
from litellm.proxy.guardrails.anthropic_sse import (
anthropic_sse_chunks_from_response,
@ -52,7 +53,12 @@ from litellm.proxy.guardrails.anthropic_sse import (
model_response_text,
)
from litellm.secret_managers.main import get_secret_str
from litellm.types.guardrails import BedrockChecksConfigModel, GuardrailEventHooks
from litellm.types.guardrails import (
BedrockChecksConfigModel,
BedrockGuardrailStreamingParams,
GuardrailEventHooks,
LitellmParams,
)
from litellm.types.llms.openai import AllMessageValues, ChatCompletionUserMessage
from litellm.types.proxy.guardrails.guardrail_hooks.bedrock_guardrails import (
BedrockChecksMessage,
@ -221,9 +227,23 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
prompt_attack_threshold: float | None = 0.5,
pii_confidence_threshold: float | None = 0.5,
chunk_budget_chars: int = BEDROCK_APPLY_GUARDRAIL_CHUNK_BUDGET_CHARS,
streaming_buffer_until_moderated: bool | None = None,
streaming_sampling_rate: int | None = None,
streaming_end_of_stream_only: bool | None = None,
**kwargs,
):
self.async_handler = get_async_httpx_client(llm_provider=httpxSpecialProvider.GuardrailCallback)
self._set_streaming_params(
BedrockGuardrailStreamingParams.from_extras(
MappingProxyType(
{
"streaming_buffer_until_moderated": streaming_buffer_until_moderated,
"streaming_sampling_rate": streaming_sampling_rate,
"streaming_end_of_stream_only": streaming_end_of_stream_only,
}
)
)
)
self.guardrailIdentifier = guardrailIdentifier
self.guardrailVersion = guardrailVersion
self.guardrail_provider = "bedrock"
@ -278,6 +298,18 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
list(self.checks.keys()) if self.checks else None,
)
def _set_streaming_params(self, streaming_params: BedrockGuardrailStreamingParams) -> None:
self.streaming_buffer_until_moderated = streaming_params.streaming_buffer_until_moderated
self.streaming_sampling_rate = streaming_params.streaming_sampling_rate
self.streaming_end_of_stream_only = streaming_params.streaming_end_of_stream_only
def update_in_memory_litellm_params(self, litellm_params: LitellmParams) -> None:
super().update_in_memory_litellm_params(litellm_params)
self._set_streaming_params(BedrockGuardrailStreamingParams.from_extras(litellm_params.model_extra))
def _streams_incrementally(self) -> bool:
return not self.streaming_buffer_until_moderated and not self.mask_response_content
@classmethod
def get_supported_event_hooks(cls) -> list[GuardrailEventHooks]:
return [
@ -2660,6 +2692,21 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
Collect content from the stream and run the bedrock OUTPUT scan
(post_call only validates the response).
"""
if self._streams_incrementally():
from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import (
UnifiedLLMGuardrails,
)
async for streamed_chunk in UnifiedLLMGuardrails().async_post_call_streaming_iterator_hook(
user_api_key_dict=user_api_key_dict,
response=response,
request_data=request_data,
guardrail_to_apply=self,
buffer_until_moderated_default=False,
):
yield streamed_chunk
return
# Import here to avoid circular imports
from litellm.llms.base_llm.base_model_iterator import MockResponseIterator
from litellm.main import stream_chunk_builder
@ -2716,7 +2763,7 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
)
if not raw_sse or (not is_block and not headers_flushed):
raise
block_message, _ = _serialize_http_exception_detail(block_detail)
block_message, _ = serialize_http_exception_detail(block_detail)
for error_frame in anthropic_sse_error_frames(
block_message if is_block else f"{block_exc.status_code}: {block_message}"
):

27
litellm/proxy/guardrails/guardrail_hooks/hiddenlayer/hiddenlayer.py
View file

@ -156,6 +156,31 @@ def _header_value(headers: Mapping[str, str], key: str, default: str) -> str:
return headers.get(key, default)
def _is_image_part(item: object) -> bool:
"""Whether a structured-message content part carries an image rather than text."""
if not isinstance(item, Mapping):
return False
part: Final[Mapping[object, object]] = item
return part.get("type") == "image_url"
def _scannable_text(content: object) -> str:
"""Flatten a structured message's content into the single string the v1 detection endpoint takes.
Image parts are dropped: the endpoint accepts one string, so an image would only reach it as
its stringified source (a base64 blob or a URL), which is not text the scanner can evaluate.
"""
if not isinstance(content, list):
return str(content or "")
parts: Final[Sequence[object]] = content
text_parts: Final = [item for item in parts if not _is_image_part(item)] # mutable-ok: sent as a list repr
return str(text_parts or "")
def is_saas(host: str) -> bool:
"""Checks whether the connection is to the SaaS platform"""
@ -270,7 +295,7 @@ class HiddenlayerGuardrail(CustomGuardrail):
"messages": [
{
"role": last_msg.get("role", "user"),
"content": str(last_msg.get("content", "")),
"content": _scannable_text(last_msg.get("content")),
}
]
},

1
litellm/proxy/guardrails/guardrail_hooks/prompt_security/__init__.py
View file

@ -20,6 +20,7 @@ def initialize_guardrail(litellm_params: "LitellmParams", guardrail: "Guardrail"
guardrail_name=guardrail.get("guardrail_name", ""),
event_hook=litellm_params.mode,
default_on=litellm_params.default_on,
file_sanitization_fail_open=getattr(litellm_params, "file_sanitization_fail_open", None),
)
litellm.logging_callback_manager.add_litellm_callback(_prompt_security_callback)

49
litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py
View file

@ -4,10 +4,12 @@ import os
from collections.abc import Mapping, Sequence
from typing import TYPE_CHECKING, Final, Literal, Optional
import httpx
from fastapi import HTTPException
from typing_extensions import ReadOnly, TypedDict
from litellm._logging import verbose_proxy_logger
from litellm.exceptions import Timeout as LiteLLMTimeout
from litellm.integrations.custom_guardrail import (
CustomGuardrail,
log_guardrail_information,
@ -24,6 +26,9 @@ if TYPE_CHECKING:
from litellm.types.proxy.guardrails.guardrail_hooks.base import GuardrailConfigModel
_SANITIZE_FILE_FAIL_OPEN_TIMEOUT_SECONDS: Final = 30.0
class PromptSecurityGuardrailMissingSecrets(Exception):
pass
@ -63,6 +68,13 @@ class _SanitizeStatusResponse(TypedDict, total=False):
metadata: ReadOnly[_SanitizeMetadata]
class _SanitizeResult(TypedDict):
action: ReadOnly[str]
content: ReadOnly[str | None]
metadata: ReadOnly[_SanitizeMetadata]
violations: ReadOnly[Sequence[str]]
class PromptSecurityGuardrail(CustomGuardrail):
@classmethod
def get_supported_event_hooks(cls) -> list[GuardrailEventHooks]:
@ -79,6 +91,8 @@ class PromptSecurityGuardrail(CustomGuardrail):
user: str | None = None,
system_prompt: str | None = None,
check_tool_results: bool | None = None,
file_sanitization_timeout: float = _SANITIZE_FILE_FAIL_OPEN_TIMEOUT_SECONDS,
file_sanitization_fail_open: bool | None = None,
**kwargs,
):
kwargs.setdefault("supported_event_hooks", list(self.get_supported_event_hooks()))
@ -108,6 +122,8 @@ class PromptSecurityGuardrail(CustomGuardrail):
# Configuration for file sanitization
self.max_poll_attempts = 30 # Maximum number of polling attempts
self.poll_interval = 2 # Seconds between polling attempts
self.file_sanitization_timeout = file_sanitization_timeout
self.file_sanitization_fail_open = file_sanitization_fail_open is not False
super().__init__(**kwargs)
@ -397,6 +413,39 @@ class PromptSecurityGuardrail(CustomGuardrail):
Sanitize file content using Prompt Security API.
Returns: dict with keys 'action', 'content', 'metadata'
"""
try:
return await asyncio.wait_for(
self._sanitize_file_content(file_data, filename, user_api_key_alias),
timeout=self.file_sanitization_timeout,
)
except (asyncio.TimeoutError, httpx.TimeoutException, LiteLLMTimeout) as exc:
if not self.file_sanitization_fail_open:
verbose_proxy_logger.error(
"Prompt Security Guardrail: file sanitization for %s timed out with %s; failing closed",
filename,
type(exc).__name__,
)
raise HTTPException(status_code=408, detail="File sanitization timeout") from exc
verbose_proxy_logger.error(
"Prompt Security Guardrail: file sanitization for %s timed out with %s; failing open",
filename,
type(exc).__name__,
)
fail_open_result: Final[_SanitizeResult] = {
"action": "allow",
"content": None,
"metadata": {},
"violations": (),
}
return fail_open_result
async def _sanitize_file_content(
self,
file_data: bytes,
filename: str,
user_api_key_alias: str | None,
) -> _SanitizeResult:
headers: Final = {"APP-ID": self.api_key}
if user_api_key_alias:
headers["X-LiteLLM-Key-Alias"] = user_api_key_alias

67
litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
View file

@ -57,6 +57,9 @@ class _EndpointTranslation(Protocol):
@property
def build_block_sse_chunks(self) -> "Callable[..., Sequence[bytes] | None]": ...
@property
def build_stream_error_items(self) -> "Callable[..., Sequence[object] | None]": ...
def _as_endpoint_translation(translation: _EndpointTranslation) -> _EndpointTranslation:
return translation
@ -408,14 +411,32 @@ class UnifiedLLMGuardrails(CustomLogger):
call_type: str | None,
responses_so_far: Sequence[object],
request_data: dict,
endpoint_translation: _EndpointTranslation | None = None,
stream_started: bool = False,
responses_yielded: Sequence[object] | None = None,
) -> AsyncGenerator[object, None]:
"""Surface a mid-stream HTTPException. For A2A call types the response has
already started, so emit an in-stream JSON-RPC error chunk; otherwise
re-raise so the proxy can report it.
"""Surface a mid-stream HTTPException (a guardrail block with the default
exception-on-block config, or a failed scan).
A2A call types emit an in-stream JSON-RPC error chunk. For other call
types, once chunks have already reached the client the HTTP status is
gone, so the failure is delegated to the endpoint translation's
``build_stream_error_items`` and travels as an in-stream error frame in
that endpoint's wire format. Before the first chunk (or when the format
has no in-stream error frame) the exception is re-raised so the proxy
can report it with a real HTTP status.
"""
if call_type is not None and CallTypes(call_type) in A2A_CALL_TYPES:
yield _a2a_jsonrpc_error_chunk(exc, _get_a2a_request_id(responses_so_far, request_data))
return
if stream_started and endpoint_translation is not None:
error_items: Final = endpoint_translation.build_stream_error_items(
exc, responses_so_far=tuple(responses_yielded) if responses_yielded is not None else None
)
if error_items is not None:
for error_item in error_items:
yield error_item
return
raise exc
def _build_transform_chunk(
@ -586,7 +607,15 @@ class UnifiedLLMGuardrails(CustomLogger):
yield block_chunk
raise _StreamTerminated()
except HTTPException as e:
async for error_item in self._emit_streaming_http_error(e, call_type, responses_so_far, request_data):
async for error_item in self._emit_streaming_http_error(
e,
call_type,
responses_so_far,
request_data,
endpoint_translation=endpoint_translation,
stream_started=bool(responses_yielded),
responses_yielded=responses_yielded,
):
yield error_item
raise _StreamTerminated()
@ -1070,11 +1099,17 @@ class UnifiedLLMGuardrails(CustomLogger):
return
except HTTPException as e:
# Response already started (we already yielded chunks); cannot send 400.
# For A2A, yield an in-stream JSON-RPC error so the client sees it.
if call_type is not None and CallTypes(call_type) in A2A_CALL_TYPES:
yield _a2a_jsonrpc_error_chunk(e, _get_a2a_request_id(responses_so_far, request_data))
return
raise
async for error_item in self._emit_streaming_http_error(
e,
call_type,
responses_so_far,
request_data,
endpoint_translation=endpoint_translation,
stream_started=chunks_yielded,
responses_yielded=responses_yielded,
):
yield error_item
return
chunks_yielded = True
responses_yielded.append(original_item)
yield original_item
@ -1133,7 +1168,13 @@ class UnifiedLLMGuardrails(CustomLogger):
yield block_chunk
return
except HTTPException as e:
if call_type is not None and CallTypes(call_type) in A2A_CALL_TYPES:
yield _a2a_jsonrpc_error_chunk(e, _get_a2a_request_id(responses_so_far, request_data))
else:
raise
async for error_item in self._emit_streaming_http_error(
e,
call_type,
responses_so_far,
request_data,
endpoint_translation=endpoint_translation,
stream_started=bool(responses_yielded),
responses_yielded=responses_yielded,
):
yield error_item

4
litellm/proxy/guardrails/guardrail_initializers.py
View file

@ -11,6 +11,7 @@ def initialize_bedrock(litellm_params: LitellmParams, guardrail: Guardrail):
BedrockGuardrail,
)
streaming_params: Final = BedrockGuardrailStreamingParams.from_extras(litellm_params.model_extra)
_bedrock_callback: Final = BedrockGuardrail(
guardrail_name=guardrail.get("guardrail_name", ""),
event_hook=litellm_params.mode,
@ -38,6 +39,9 @@ def initialize_bedrock(litellm_params: LitellmParams, guardrail: Guardrail):
aws_bedrock_runtime_endpoint=litellm_params.aws_bedrock_runtime_endpoint,
experimental_use_latest_role_message_only=litellm_params.experimental_use_latest_role_message_only,
only_scan_new_messages=litellm_params.only_scan_new_messages or False,
streaming_buffer_until_moderated=streaming_params.streaming_buffer_until_moderated,
streaming_sampling_rate=streaming_params.streaming_sampling_rate,
streaming_end_of_stream_only=streaming_params.streaming_end_of_stream_only,
)
litellm.logging_callback_manager.add_litellm_callback(_bedrock_callback)
return _bedrock_callback

6
litellm/proxy/management_endpoints/customer_endpoints.py
View file

@ -626,11 +626,7 @@ async def update_end_user(
# get non default values for key
non_default_values: Final = dict[str, object]()
for k, v in data_json.items():
if v is not None and v not in (
[],
{},
0,
): # models default to [], spend defaults to 0, we should not reset these values
if v is not None and ((isinstance(v, bool) and k in data.fields_set()) or v not in ([], {}, 0)):
non_default_values[k] = v
## Get end user table data ##

103
litellm/proxy/management_endpoints/internal_user_endpoints.py
View file

@ -17,7 +17,7 @@ import json
import traceback
from collections.abc import Awaitable, Mapping, Sequence
from datetime import datetime, timezone
from typing import Any, Final, Literal, cast
from typing import Any, Final, Literal, Protocol, cast, overload
import fastapi
from fastapi import APIRouter, Depends, Header, HTTPException, Request, status
@ -735,10 +735,44 @@ def _enforce_user_info_access(user_id: str | None, user_api_key_dict: UserAPIKey
)
async def _get_user_info_teams(
prisma_client: Any,
class _UserInfoDataClient(Protocol):
@overload
async def get_data(self, *, user_id: str) -> "prisma_models.LiteLLM_UserTable | None": ...
@overload
async def get_data(
self,
*,
user_id: str | None,
table_name: Literal["key"],
query_type: Literal["find_all"],
) -> "Sequence[LiteLLM_VerificationToken] | None": ...
@overload
async def get_data(
self,
*,
team_id_list: list[str],
table_name: Literal["team"],
query_type: Literal["find_all"],
) -> "Sequence[TeamListResponseObject] | None": ...
async def _get_user_info_keys(
prisma_client: "_UserInfoDataClient",
user_id: str | None,
user_info: Any | None,
) -> "Sequence[LiteLLM_VerificationToken] | None":
return await prisma_client.get_data(
user_id=user_id,
table_name="key",
query_type="find_all",
)
async def _get_user_info_teams(
prisma_client: "_UserInfoDataClient",
user_id: str | None,
user_info: "prisma_models.LiteLLM_UserTable",
user_api_key_dict: UserAPIKeyAuth,
) -> tuple[list[TeamListResponseObject], list[TeamListResponseObject] | None]:
"""Fetch and merge teams from membership + user.teams field."""
@ -759,7 +793,7 @@ async def _get_user_info_teams(
team_list = teams_1
team_id_list = [team.team_id for team in teams_1]
teams_2: list[TeamListResponseObject] | None = None
teams_2: Sequence[TeamListResponseObject] | None = None
target_team_ids: Final = getattr(user_info, "teams", None)
if target_team_ids and isinstance(target_team_ids, list):
@ -769,8 +803,8 @@ async def _get_user_info_teams(
query_type="find_all",
)
elif user_api_key_dict.user_id is not None and user_id is None:
caller_user_info: Final[object] = await prisma_client.get_data(user_id=user_api_key_dict.user_id)
caller_team_ids: Final = getattr(caller_user_info, "teams", None)
caller_user_info: Final = await prisma_client.get_data(user_id=user_api_key_dict.user_id)
caller_team_ids: Final = caller_user_info.teams if caller_user_info is not None else None
if caller_team_ids:
teams_2 = await prisma_client.get_data(
team_id_list=caller_team_ids,
@ -807,7 +841,7 @@ def _redact_scim_enterprise_metadata(
def _build_user_info_response(
user_id: str | None,
user_info: Any | None,
keys: list[LiteLLM_VerificationToken] | None,
keys: Sequence[LiteLLM_VerificationToken] | None,
team_list: list[TeamListResponseObject],
teams_1: list[TeamListResponseObject] | None,
model_max_budget_usage: dict[str, dict[str, object]] | None = None,
@ -894,11 +928,7 @@ async def user_info(
)
## GET ALL KEYS ##
keys: Final = await prisma_client.get_data(
user_id=user_id,
table_name="key",
query_type="find_all",
)
keys: Final = await _get_user_info_keys(prisma_client, user_id)
response_data: Final = _build_user_info_response(
user_id=user_id,
@ -997,6 +1027,14 @@ async def user_info_v2(
This is the v2 replacement for /user/info, designed to avoid the "god endpoint" problem
where the old endpoint loaded all keys and teams into memory.
Note on `spend`: this is the user's running budget counter, which the budget reset job
resets whenever `budget_reset_at` elapses (see `budget_duration`): to zero by default,
or to the overage above `max_budget` when `budget_rollover` is enabled. It is NOT
lifetime or per-period historical spend. For historical spend over a date range, use
`/user/daily/activity` or `/user/daily/activity/aggregated`, which read daily spend
records that only ever accumulate and are never reset. The two values are expected to
diverge once a budget reset has occurred within the queried period.
Access control:
- Proxy admins can query any user
- Team admins can query users within their teams
@ -1077,6 +1115,12 @@ async def user_info_v2(
raise handle_exception_on_proxy(e)
async def _fetch_admin_teams_and_keys_rows(
prisma_client: "PrismaClient", sql_query: str
) -> Sequence[Mapping[str, Sequence[Mapping[str, object]] | None]]:
return await prisma_client.db.query_raw(sql_query)
async def _get_user_info_for_proxy_admin(user_api_key_dict: UserAPIKeyAuth):
"""
Admin UI Endpoint - Returns All Teams and Keys when Proxy Admin is querying
@ -1100,22 +1144,25 @@ async def _get_user_info_for_proxy_admin(user_api_key_dict: UserAPIKeyAuth):
"Database not connected. Connect a database to your proxy - https://docs.litellm.ai/docs/simple_proxy#managing-auth---virtual-keys"
)
results: Final = await prisma_client.db.query_raw(sql_query)
results: Final = await _fetch_admin_teams_and_keys_rows(prisma_client, sql_query)
verbose_proxy_logger.debug("results_keys: %s", results)
_keys_in_db: Final[Sequence[dict[str, object]]] = results[0]["keys"] or []
_keys_in_db: Final[Sequence[Mapping[str, object]]] = results[0]["keys"] or []
# cast all keys to LiteLLM_VerificationToken
keys_in_db: Final = []
for key in _keys_in_db:
if key.get("models") is None:
key["models"] = []
keys_in_db.append(LiteLLM_VerificationToken.model_validate(key))
key_payload = dict[str, object](key)
if key_payload.get("models") is None:
key_payload["models"] = []
keys_in_db.append(LiteLLM_VerificationToken.model_validate(key_payload))
# cast all teams to LiteLLM_TeamTable
_teams_in_db: list[LiteLLM_TeamTable] = results[0]["teams"] or []
_teams_in_db = [LiteLLM_TeamTable.model_validate(team) for team in _teams_in_db]
_teams_in_db.sort(key=lambda x: getattr(x, "team_alias", "") or "")
_teams_rows: Final[Sequence[Mapping[str, object]]] = results[0]["teams"] or []
_teams_in_db: Final = sorted(
(LiteLLM_TeamTable.model_validate(team) for team in _teams_rows),
key=lambda x: getattr(x, "team_alias", "") or "",
)
returned_keys: Final = _process_keys_for_user_info(keys=keys_in_db, all_teams=_teams_in_db)
# Get admin's own user_id and user_info
@ -1140,7 +1187,7 @@ async def _get_user_info_for_proxy_admin(user_api_key_dict: UserAPIKeyAuth):
def _process_keys_for_user_info(
keys: list[LiteLLM_VerificationToken] | None,
keys: Sequence[LiteLLM_VerificationToken] | None,
all_teams: list[LiteLLM_TeamTable] | list[TeamListResponseObject] | None,
):
from litellm.constants import UI_SESSION_TOKEN_TEAM_ID
@ -1231,7 +1278,7 @@ def _update_internal_user_params(data_json: dict, data: UpdateUserRequest | Upda
async def _schedule_user_update_audit_log(
response: dict[str, Any],
response: Mapping[str, object],
existing_user_row: BaseModel | None,
litellm_changed_by: str | None,
user_api_key_dict: UserAPIKeyAuth,
@ -2687,6 +2734,11 @@ async def get_user_daily_activity(
Meant to optimize querying spend data for analytics for a user.
Reads daily spend records that only ever accumulate and are never affected by budget
resets. Their total can legitimately exceed the `spend` field returned by
`/v2/user/info`, which is a running budget counter that every budget reset sets back
to zero (or to the overage above `max_budget` when `budget_rollover` is enabled).
Returns:
(by date)
- spend
@ -2800,6 +2852,11 @@ async def get_user_daily_activity_aggregated(
"""
Aggregated analytics for a user's daily activity without pagination.
Returns the same response shape as the paginated endpoint with page metadata set to single-page.
Reads daily spend records that only ever accumulate and are never affected by budget
resets. Their total can legitimately exceed the `spend` field returned by
`/v2/user/info`, which is a running budget counter that every budget reset sets back
to zero (or to the overage above `max_budget` when `budget_rollover` is enabled).
"""
from litellm.proxy.proxy_server import prisma_client

100
litellm/proxy/management_endpoints/key_management_endpoints.py
View file

@ -18,7 +18,8 @@ import os
import re
import secrets
import traceback
from collections.abc import Awaitable, Callable, Mapping, Sequence
from collections.abc import Awaitable, Callable, Iterator, Mapping, Sequence
from contextlib import AbstractAsyncContextManager
from datetime import datetime, timedelta, timezone
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, Protocol, TypeVar, cast
@ -230,6 +231,54 @@ def _config_table(prisma_client: PrismaClient) -> _ConfigTableActions:
)
class _CustomKeyHooksModule(Protocol):
user_custom_key_generate: Callable[..., Awaitable[Mapping[str, object]]] | None
user_custom_key_update: Callable[..., Awaitable[Mapping[str, object]]] | None
def _custom_key_generate_hook(
hooks: _CustomKeyHooksModule,
) -> Callable[..., Awaitable[Mapping[str, object]]] | None:
return hooks.user_custom_key_generate
def _custom_key_update_hook(
hooks: _CustomKeyHooksModule,
) -> Callable[..., Awaitable[Mapping[str, object]]] | None:
return hooks.user_custom_key_update
class _LegacyDumpable(Protocol):
def dict(self) -> Mapping[str, object]: ...
def _legacy_model_dict(row: _LegacyDumpable) -> Mapping[str, object]:
return row.dict()
def _as_object_dict(values: Mapping[str, object]) -> Mapping[str, object]:
return values
def _model_items(model: BaseModel) -> Iterator[tuple[str, object]]:
return iter(model)
class _EnvVarsParam(Protocol):
@property
def param_value(self) -> Mapping[str, str] | None: ...
def _env_vars_param_value(param: _EnvVarsParam) -> Mapping[str, str] | None:
return param.param_value
def _tx_tables_context(
open_tx: Callable[[], AbstractAsyncContextManager[_TxTables]],
) -> AbstractAsyncContextManager[_TxTables]:
return open_tx()
async def _check_custom_key_allowed(custom_key_value: str | None) -> None:
"""Raise 403 if custom API keys are disabled and a custom key was provided."""
if custom_key_value is None:
@ -910,7 +959,7 @@ async def _common_key_generation_helper(
# check if user set default key/generate params on config.yaml
if litellm.default_key_generate_params is not None:
for elem in data:
for elem in _model_items(data):
key, value = elem
if (
value is None
@ -1692,11 +1741,11 @@ async def generate_key_fn(
- user_id: (str) Unique user id - used for tracking spend across multiple keys for same user id.
"""
try:
from litellm.proxy import proxy_server
from litellm.proxy._types import CommonProxyErrors
from litellm.proxy.proxy_server import (
prisma_client,
user_api_key_cache,
user_custom_key_generate,
)
if prisma_client is None:
@ -1723,7 +1772,7 @@ async def generate_key_fn(
)
custom_key_generate_hook: Final[Callable[..., Awaitable[Mapping[str, object]]] | None] = (
user_custom_key_generate
_custom_key_generate_hook(proxy_server)
)
if custom_key_generate_hook is not None:
if inspect.iscoroutinefunction(custom_key_generate_hook):
@ -1892,11 +1941,11 @@ async def generate_service_account_key_fn(
- user_id: (str) Unique user id - used for tracking spend across multiple keys for same user id.
"""
from litellm.proxy import proxy_server
from litellm.proxy._types import CommonProxyErrors
from litellm.proxy.proxy_server import (
prisma_client,
user_api_key_cache,
user_custom_key_generate,
)
if prisma_client is None:
@ -1924,7 +1973,9 @@ async def generate_service_account_key_fn(
verbose_proxy_logger.debug("entered /key/generate")
custom_key_generate_hook: Final[Callable[..., Awaitable[Mapping[str, object]]] | None] = user_custom_key_generate
custom_key_generate_hook: Final[Callable[..., Awaitable[Mapping[str, object]]] | None] = _custom_key_generate_hook(
proxy_server
)
if custom_key_generate_hook is not None:
if inspect.iscoroutinefunction(custom_key_generate_hook):
result: Final = await custom_key_generate_hook(data)
@ -1998,7 +2049,7 @@ def prepare_metadata_fields(data: BaseModel, non_default_values: dict, existing_
)
casted_metadata[reserved_field] = existing_value
data_json: Final[Mapping[str, object]] = data.model_dump(exclude_unset=True, exclude_none=True)
data_json: Final = _as_object_dict(data.model_dump(exclude_unset=True, exclude_none=True))
try:
for k, v in data_json.items():
@ -2805,13 +2856,13 @@ async def update_key_fn(
}'
```
"""
from litellm.proxy import proxy_server
from litellm.proxy.proxy_server import (
llm_router,
premium_user,
prisma_client,
proxy_logging_obj,
user_api_key_cache,
user_custom_key_update,
)
try:
@ -2842,7 +2893,9 @@ async def update_key_fn(
)
# Custom key update hook
custom_key_update_hook: Final[Callable[..., Awaitable[Mapping[str, object]]] | None] = user_custom_key_update
custom_key_update_hook: Final[Callable[..., Awaitable[Mapping[str, object]]] | None] = _custom_key_update_hook(
proxy_server
)
if custom_key_update_hook is not None:
if inspect.iscoroutinefunction(custom_key_update_hook):
result: Final = await custom_key_update_hook(data)
@ -3004,14 +3057,16 @@ async def bulk_update_keys(
}'
```
"""
from litellm.proxy import proxy_server
from litellm.proxy.proxy_server import (
llm_router,
prisma_client,
proxy_logging_obj,
user_api_key_cache,
user_custom_key_update,
)
custom_key_update_hook: Final = _custom_key_update_hook(proxy_server)
if user_api_key_dict.user_role != LitellmUserRoles.PROXY_ADMIN.value:
raise HTTPException(
status_code=403,
@ -3057,7 +3112,7 @@ async def bulk_update_keys(
user_api_key_cache=user_api_key_cache,
proxy_logging_obj=proxy_logging_obj,
llm_router=llm_router,
user_custom_key_update=user_custom_key_update,
user_custom_key_update=custom_key_update_hook,
)
successful_updates.append(
@ -3135,7 +3190,7 @@ def _build_failed_team_key_update(
if hasattr(existing_key_row, "model_dump"):
key_info = existing_key_row.model_dump()
elif hasattr(existing_key_row, "dict"):
key_info = existing_key_row.dict()
key_info = dict[str, object](_legacy_model_dict(existing_key_row))
if key_info:
key_info.pop("token", None)
@ -3166,14 +3221,16 @@ async def bulk_update_team_keys(
Callable by proxy admins, or by team admins with `KEY_UPDATE` permission.
"""
from litellm.proxy import proxy_server
from litellm.proxy.proxy_server import (
llm_router,
prisma_client,
proxy_logging_obj,
user_api_key_cache,
user_custom_key_update,
)
custom_key_update_hook: Final = _custom_key_update_hook(proxy_server)
if prisma_client is None:
raise HTTPException(
status_code=500,
@ -3302,7 +3359,7 @@ async def bulk_update_team_keys(
user_api_key_cache=user_api_key_cache,
proxy_logging_obj=proxy_logging_obj,
llm_router=llm_router,
user_custom_key_update=user_custom_key_update,
user_custom_key_update=custom_key_update_hook,
existing_key_row=existing_by_token[db_token],
)
@ -4301,7 +4358,7 @@ def _transform_verification_tokens_to_deleted_records(
"litellm_changed_by": litellm_changed_by,
}
)
record = deleted_record.model_dump()
record = dict[str, object](_as_object_dict(deleted_record.model_dump()))
# Map org_id to organization_id (model uses org_id, but schema expects organization_id)
org_id_value: object = record.pop("org_id", None)
@ -4437,13 +4494,12 @@ async def _rotate_master_key(
should_create_model_in_db=False,
)
if new_model:
_dumped = new_model.model_dump(exclude_none=True)
_dumped = dict[str, object](_as_object_dict(new_model.model_dump(exclude_none=True)))
_dumped["litellm_params"] = prisma.Json(_dumped["litellm_params"])
_dumped["model_info"] = prisma.Json(_dumped["model_info"])
new_models.append(_dumped)
verbose_proxy_logger.debug("Resetting proxy model table")
async with prisma_client.db.tx() as tx_ctx:
tx: Final[_TxTables] = tx_ctx
async with _tx_tables_context(prisma_client.db.tx) as tx:
await tx.litellm_proxymodeltable.delete_many()
verbose_proxy_logger.debug("Creating %s models", len(new_models))
await tx.litellm_proxymodeltable.create_many(
@ -4458,14 +4514,14 @@ async def _rotate_master_key(
if config:
"""If environment_variables is found, decrypt it and encrypt it with the new master key"""
environment_variables_dict = {}
environment_variables_dict: Mapping[str, str] | None = {}
for c in config:
if c.param_name == "environment_variables":
environment_variables_dict = c.param_value
environment_variables_dict = _env_vars_param_value(c)
if environment_variables_dict:
decrypted_env_vars: Final = proxy_config._decrypt_and_set_db_env_variables(
environment_variables=environment_variables_dict
environment_variables=dict[str, str](environment_variables_dict)
)
encrypted_env_vars: Final = proxy_config._encrypt_env_variables(
environment_variables=decrypted_env_vars,
@ -4531,7 +4587,7 @@ async def _rotate_master_key(
updated_patch=decrypted_cred,
new_encryption_key=new_master_key,
)
_cred_data = encrypted_cred.model_dump(exclude_none=True)
_cred_data = dict[str, object](_as_object_dict(encrypted_cred.model_dump(exclude_none=True)))
if "credential_values" in _cred_data:
_cred_data["credential_values"] = prisma.Json(_cred_data["credential_values"])
if "credential_info" in _cred_data:

4
litellm/proxy/management_endpoints/team_callback_endpoints.py
View file

@ -252,7 +252,7 @@ async def add_team_callbacks(
Use this if if you want different teams to have different success/failure callbacks
Parameters:
- callback_name (Literal["langfuse", "langsmith", "gcs"], required): The name of the callback to add
- callback_name (str, required): The name of the callback to add, e.g. "langfuse", "langsmith", "gcs", "newrelic". The value is validated against the callbacks that support team-scoped credentials
- callback_type (Literal["success", "failure", "success_and_failure"], required): The type of callback to add. One of:
- "success": Callback for successful LLM calls
- "failure": Callback for failed LLM calls
@ -268,6 +268,8 @@ async def add_team_callbacks(
- langsmith_api_key: The API key for the Langsmith callback
- langsmith_project: The project for the Langsmith callback
- langsmith_base_url: The base URL for the Langsmith callback
- newrelic_api_key: The ingest license key for the team's New Relic account; routes both LLM/agent traces and cost metrics to that account. Requires the proxy to run with LITELLM_OTEL_V2=true, otherwise this callback is rejected with a 400
- newrelic_region: The New Relic region for the team's account ("us" or "eu"), riding the team's own key
Example curl:
```

41
litellm/proxy/pass_through_endpoints/streaming_handler.py
View file

@ -65,6 +65,19 @@ class PassThroughStreamingHandler:
route_streaming_logging or PassThroughStreamingHandler._route_streaming_logging_to_handler
)
raw_bytes: Final[list[bytes]] = []
def _build_logging_coroutine() -> Coroutine[None, None, None]:
return resolved_route_streaming_logging(
litellm_logging_obj=litellm_logging_obj,
passthrough_success_handler_obj=passthrough_success_handler_obj,
url_route=url_route,
request_body=request_body or {},
endpoint_type=endpoint_type,
start_time=start_time,
raw_bytes=raw_bytes,
end_time=datetime.now(),
)
logging_scheduled = False
model_name: Final = PassThroughStreamingHandler._extract_model_for_cost_injection(
request_body=request_body,
@ -114,6 +127,21 @@ class PassThroughStreamingHandler:
)
if pending:
yield pending
# Stream completed cleanly. When the proxy armed deferred
# dispatch (post-call guardrails active), park the logging
# coroutine on logging_obj instead of enqueueing now, so
# ProxyLogging._fire_deferred_stream_logging fires it after
# guardrail end-of-stream blocks populate guardrail_information.
# Disconnect/exception paths skip this and fall through to the
# immediate enqueue in ``finally`` to keep partial billing
# (LIT-2642).
if (
getattr(litellm_logging_obj, "_on_deferred_stream_complete", None) is not None
and raw_bytes
and response.status_code < 400
):
logging_scheduled = True
litellm_logging_obj._deferred_stream_complete_args = (_build_logging_coroutine(),)
except Exception as e:
verbose_proxy_logger.error("Error in chunk_processor: %s", e)
raise
@ -128,18 +156,7 @@ class PassThroughStreamingHandler:
if not logging_scheduled and raw_bytes and response.status_code < 400:
logging_scheduled = True
try:
GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(
async_coroutine=resolved_route_streaming_logging(
litellm_logging_obj=litellm_logging_obj,
passthrough_success_handler_obj=passthrough_success_handler_obj,
url_route=url_route,
request_body=request_body or {},
endpoint_type=endpoint_type,
start_time=start_time,
raw_bytes=raw_bytes,
end_time=datetime.now(),
)
)
GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(async_coroutine=_build_logging_coroutine())
except Exception as e:
verbose_proxy_logger.error("Error scheduling chunk_processor logging: %s", e)

4
litellm/proxy/prompts/prompt_endpoints.py
View file

@ -6,7 +6,7 @@ import tempfile
from collections.abc import Awaitable, Mapping, Sequence
from datetime import datetime
from pathlib import Path
from typing import TYPE_CHECKING, Any, Final, Protocol, cast
from typing import TYPE_CHECKING, Final, Protocol, cast
from fastapi import (
APIRouter,
@ -1317,7 +1317,7 @@ async def test_prompt(
async def convert_prompt_file_to_json(
file: UploadFile = File(...),
user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth),
) -> dict[str, Any]:
) -> Mapping[str, object]:
"""
Convert a .prompt file to JSON format.

3
litellm/proxy/proxy_server.py
View file

@ -1363,7 +1363,7 @@ _OPENAPI_HTTP_METHODS: Final = {
# the UI. Kept here at module scope to match the analogous descriptor
# `is_secret` flags in litellm.proxy.config_resolvers and the
# `_CACHE_SENSITIVE_FIELDS` constant in the cache endpoint file.
_ALERTING_SENSITIVE_VARS: Final[set[str]] = {"SLACK_WEBHOOK_URL", "SMTP_PASSWORD"}
_ALERTING_SENSITIVE_VARS: Final[set[str]] = {"ALERTING_WEBHOOK_URL", "SLACK_WEBHOOK_URL", "SMTP_PASSWORD"}
def _strip_operation_id_method_suffix(operation_id: str) -> str:
@ -16566,6 +16566,7 @@ async def create_config_audit_log(
_EXTRA_SECRET_CALLBACK_ENV_VARS: Final = frozenset(
{
"ALERTING_WEBHOOK_URL",
"GALILEO_USERNAME",
"GENERIC_LOGGER_HEADERS",
"OTEL_HEADERS",

163
litellm/proxy/response_api_endpoints/endpoints.py
View file

@ -1,14 +1,18 @@
import asyncio
import json
import time
from collections.abc import AsyncIterator, Mapping
from collections.abc import AsyncIterator, Awaitable, Mapping
from enum import Enum
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, NamedTuple, cast, get_args
from typing import TYPE_CHECKING, Any, Final, NamedTuple, Protocol, cast, get_args
from uuid import uuid4
import fastapi
from fastapi import APIRouter, Depends, HTTPException, Request, Response
from fastapi.responses import JSONResponse
from openai.types.responses.response_create_params import ResponseInputParam
from starlette.websockets import WebSocket, WebSocketDisconnect
from typing_extensions import ReadOnly, TypedDict
from litellm._logging import verbose_proxy_logger
from litellm.integrations.custom_guardrail import ModifyResponseException
@ -26,8 +30,13 @@ from litellm.proxy.common_utils.http_parsing_utils import (
_read_request_body,
_safe_set_request_parsed_body,
)
from litellm.types.llms.openai import REASONING_EFFORT, ResponsesAPIResponse
from litellm.types.llms.openai import (
REASONING_EFFORT,
ResponsesAPIOptionalRequestParams,
ResponsesAPIResponse,
)
from litellm.types.responses.main import DeleteResponseResult
from litellm.types.utils import TokenCountResponse
if TYPE_CHECKING:
from litellm.router import Router
@ -35,7 +44,7 @@ if TYPE_CHECKING:
router: Final = APIRouter()
_user_api_key_auth_dep: Final = Depends(user_api_key_auth)
_RESPONSES_TAGS: Final = ["responses"] # mutable-ok: fastapi's route signature requires List[str] tags
_RESPONSES_TAGS: Final[list[str | Enum]] = ["responses"] # mutable-ok: fastapi's route signature requires list tags
_TOOL_PAYLOAD_KEYS: Final[Mapping[str, tuple[str, ...]]] = MappingProxyType(
{
@ -1017,6 +1026,152 @@ async def compact_response(
)
class _ResponsesApiErrorDetail(TypedDict):
message: ReadOnly[str]
type: ReadOnly[str]
param: ReadOnly[str | None]
code: ReadOnly[str | None]
class _ResponsesApiErrorBody(TypedDict):
error: ReadOnly[_ResponsesApiErrorDetail]
class _ResponsesInputTokensResult(TypedDict):
object: ReadOnly[str]
input_tokens: ReadOnly[int]
class _TokenCountPayload(TypedDict):
model: ReadOnly[str]
messages: ReadOnly[tuple[Mapping[str, object], ...]]
tools: ReadOnly[object]
class _TokenCounter(Protocol):
def __call__(self, request: TokenCountRequest, call_endpoint: bool) -> Awaitable[TokenCountResponse]: ...
def _proxy_token_counter() -> _TokenCounter:
from litellm.proxy.proxy_server import token_counter
return token_counter
_token_counter_dep: Final = Depends(_proxy_token_counter)
def _responses_invalid_request_response(message: str, param: str | None, code: str | None) -> JSONResponse:
body: Final[_ResponsesApiErrorBody] = {
"error": {
"message": message,
"type": "invalid_request_error",
"param": param,
"code": code,
}
}
return JSONResponse(status_code=400, content=body)
def _missing_responses_param_response(param: str) -> JSONResponse:
return _responses_invalid_request_response(
message=f"Missing required parameter: '{param}'.",
param=param,
code="missing_required_parameter",
)
def _responses_input_as_token_count_messages(
input_value: str | ResponseInputParam,
instructions: str | None,
) -> tuple[Mapping[str, object], ...]:
from litellm.responses.litellm_completion_transformation.transformation import (
LiteLLMCompletionResponsesConfig,
)
request_params: Final[ResponsesAPIOptionalRequestParams] = {"instructions": instructions}
transformed: Final = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
input=input_value,
responses_api_request=request_params,
)
return tuple(
message if isinstance(message, dict) else message.model_dump(exclude_none=True) for message in transformed
)
@router.post(
"/v1/responses/input_tokens",
dependencies=(_user_api_key_auth_dep,),
tags=_RESPONSES_TAGS,
)
@router.post(
"/responses/input_tokens",
dependencies=(_user_api_key_auth_dep,),
tags=_RESPONSES_TAGS,
)
@router.post(
"/openai/v1/responses/input_tokens",
dependencies=(_user_api_key_auth_dep,),
tags=_RESPONSES_TAGS,
)
async def responses_input_tokens(
request: Request,
token_counter: _TokenCounter = _token_counter_dep,
):
"""
Count the input tokens of a Responses API request without calling the model.
Follows the OpenAI Responses API spec: https://platform.openai.com/docs/api-reference/responses/input-tokens
```bash
curl -X POST http://localhost:4000/v1/responses/input_tokens \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-1234" \
-d '{
"model": "gpt-4o",
"input": "Hello, how are you?"
}'
```
Returns: `{"object": "response.input_tokens", "input_tokens": <count>}`
"""
data: Final = await _read_request_body(request=request)
model_name: Final = data.get("model")
input_value: Final = data.get("input")
if not isinstance(model_name, str) or not model_name:
return _missing_responses_param_response("model")
if input_value is None:
return _missing_responses_param_response("input")
if isinstance(input_value, (str, list)) and not input_value:
return _responses_invalid_request_response(
message="""One of "input" or "previous_response_id" or 'prompt' or 'conversation' must be provided.""",
param=None,
code="missing_required_parameter",
)
try:
payload: Final[_TokenCountPayload] = {
"model": model_name,
"messages": _responses_input_as_token_count_messages(
input_value=input_value,
instructions=data.get("instructions"),
),
"tools": data.get("tools"),
}
token_request: Final = TokenCountRequest.model_validate(payload)
except Exception as e:
return _responses_invalid_request_response(
message=f"Invalid request for token counting: {e}", param=None, code=None
)
token_response: Final = await token_counter(request=token_request, call_endpoint=True)
result: Final[_ResponsesInputTokensResult] = {
"object": "response.input_tokens",
"input_tokens": token_response.total_tokens,
}
return result
@router.post(
"/v1/responses/{response_id}/cancel",
dependencies=[Depends(user_api_key_auth)],

110
litellm/proxy/response_polling/background_streaming.py
View file

@ -10,12 +10,12 @@ https://platform.openai.com/docs/api-reference/responses-streaming
import asyncio
import json
from collections.abc import Mapping, Sequence
from typing import TYPE_CHECKING, Final, TypedDict
from collections.abc import Callable, Mapping, Sequence
from typing import TYPE_CHECKING, Final, TypeAlias
from fastapi import Request, Response
from fastapi.responses import StreamingResponse
from typing_extensions import ReadOnly
from typing_extensions import ReadOnly, TypedDict
from litellm._logging import verbose_proxy_logger
from litellm.proxy.auth.user_api_key_auth import UserAPIKeyAuth
@ -29,53 +29,59 @@ if TYPE_CHECKING:
from litellm.router import Router
class _StreamContentPart(TypedDict, total=False):
text: ReadOnly[str]
_JsonDict: TypeAlias = dict[str, object]
_JsonList: TypeAlias = list[object]
class _StreamOutputItem(TypedDict, total=False):
class _OutputItem(TypedDict, total=False):
id: ReadOnly[str]
content: ReadOnly[Sequence[_StreamContentPart | None]]
content: ReadOnly[Sequence[object]]
class _StreamTerminalResponse(TypedDict, total=False):
"""Fields of the ``response`` payload carried by a terminal streaming event."""
class _TerminalResponse(TypedDict, total=False):
status: ReadOnly[ResponsesAPIStatus]
error: ReadOnly[_JsonDict]
usage: ReadOnly[_JsonDict]
reasoning: ReadOnly[_JsonDict]
tool_choice: ReadOnly[object]
tools: ReadOnly[_JsonList]
model: ReadOnly[str]
instructions: ReadOnly[str]
temperature: ReadOnly[float]
top_p: ReadOnly[float]
max_output_tokens: ReadOnly[int]
previous_response_id: ReadOnly[str]
text: ReadOnly[_JsonDict]
truncation: ReadOnly[str]
parallel_tool_calls: ReadOnly[bool]
user: ReadOnly[str]
store: ReadOnly[bool]
output: ReadOnly[Sequence[_StreamOutputItem]]
incomplete_details: ReadOnly[_JsonDict]
output: ReadOnly[Sequence[_OutputItem]]
class _StreamEvent(TypedDict, total=False):
"""One decoded ``data:`` frame of an OpenAI Responses streaming body."""
type: ReadOnly[str]
item: ReadOnly[_StreamOutputItem]
item: ReadOnly[_OutputItem]
item_id: ReadOnly[str]
part: ReadOnly[_StreamContentPart]
content_index: ReadOnly[int]
delta: ReadOnly[str]
response: ReadOnly[_StreamTerminalResponse]
part: ReadOnly[object]
response: ReadOnly[_TerminalResponse]
class _StreamEventParser:
parse: Callable[[str], _StreamEvent] = staticmethod(json.loads)
async def background_streaming_task(
polling_id: str,
data,
data: dict,
polling_handler: ResponsePollingHandler,
request: Request,
fastapi_response: Response,
user_api_key_dict: UserAPIKeyAuth,
general_settings,
general_settings: dict,
llm_router: "Router | None",
proxy_config: "ProxyConfig",
proxy_logging_obj: "ProxyLogging",
@ -138,9 +144,10 @@ async def background_streaming_task(
# Process streaming response following OpenAI events format
# https://platform.openai.com/docs/api-reference/responses-streaming
output_items: Final[dict[str, _StreamOutputItem]] = {} # Track output items by ID
# Track accumulated text deltas by (item_id, content_index)
accumulated_text: Final[dict[tuple[str, int], str]] = {}
output_items: Final = dict[str, _OutputItem]() # Track output items by ID
accumulated_text: Final = dict[
tuple[str, int], str
]() # Track accumulated text deltas by (item_id, content_index)
# ResponsesAPIResponse fields to extract from response.completed
usage_data = None
@ -210,7 +217,7 @@ async def background_streaming_task(
break
try:
event: _StreamEvent = json.loads(chunk_data)
event: _StreamEvent = _StreamEventParser.parse(chunk_data)
event_type = event.get("type", "")
# Process different event types based on OpenAI streaming spec
@ -229,19 +236,18 @@ async def background_streaming_task(
if item_id and item_id in output_items:
# Update the output item with new content
current_item = output_items[item_id]
appended_item: _StreamOutputItem = {
**current_item,
"content": (*current_item.get("content", ()), content_part),
added_item = output_items[item_id]
output_items[item_id] = {
**added_item,
"content": (*added_item.get("content", ()), content_part),
}
output_items[item_id] = appended_item
state_dirty = True
elif event_type == "response.output_text.delta":
# Text delta - accumulate text content
# https://platform.openai.com/docs/api-reference/responses-streaming/response-text-delta
item_id = event.get("item_id")
content_index: int = event.get("content_index", 0)
content_index = event.get("content_index", 0)
delta = event.get("delta", "")
if item_id and item_id in output_items:
@ -252,24 +258,14 @@ async def background_streaming_task(
accumulated_text[key] += delta
# Update the content in output_items
current_item = output_items[item_id]
content_list: Sequence[_StreamContentPart | None] = current_item.get("content", ())
if content_index < len(content_list):
# Update existing content part with accumulated text
content_entry = content_list[content_index]
if isinstance(content_entry, dict):
delta_part: _StreamContentPart = {
**content_entry,
"text": accumulated_text[key],
}
delta_item: _StreamOutputItem = {
**current_item,
"content": tuple(
delta_part if index == content_index else entry
for index, entry in enumerate(content_list)
),
}
output_items[item_id] = delta_item
delta_item = output_items[item_id]
if "content" in delta_item:
content_list = delta_item["content"]
if content_index < len(content_list):
# Update existing content part with accumulated text
content_entry = content_list[content_index]
if isinstance(content_entry, dict):
content_entry["text"] = accumulated_text[key]
state_dirty = True
elif event_type == "response.content_part.done":
@ -280,17 +276,17 @@ async def background_streaming_task(
if item_id and item_id in output_items:
# Update with final content from event
current_item = output_items[item_id]
content_list = current_item.get("content", ())
if content_index < len(content_list):
finalized_item: _StreamOutputItem = {
**current_item,
"content": tuple(
content_part if index == content_index else entry
for index, entry in enumerate(content_list)
),
}
output_items[item_id] = finalized_item
done_item = output_items[item_id]
if "content" in done_item:
content_list = done_item["content"]
if content_index < len(content_list):
output_items[item_id] = {
**done_item,
"content": tuple(
content_part if part_index == content_index else existing_part
for part_index, existing_part in enumerate(content_list)
),
}
state_dirty = True
elif event_type == "response.output_item.done":

9
litellm/proxy/spend_tracking/budget_reservation.py
View file

@ -172,7 +172,14 @@ async def reserve_budget_for_request(
) -> dict | None:
if valid_token is None or not RouteChecks.is_llm_api_route(route=route):
return None
if route in {"/models", "/v1/models", "/utils/token_counter"}:
if route in {
"/models",
"/v1/models",
"/utils/token_counter",
"/responses/input_tokens",
"/v1/responses/input_tokens",
"/openai/v1/responses/input_tokens",
}:
return None
if get_model_from_request(request_body, route, llm_router=llm_router) is None:
return None

60
litellm/proxy/spend_tracking/spend_tracking_utils.py
View file

@ -30,7 +30,7 @@ from litellm.litellm_core_utils.litellm_logging import (
request_model_access_groups_from_litellm_params,
)
from litellm.litellm_core_utils.safe_json_dumps import safe_dumps, strip_null_bytes
from litellm.proxy._types import SpendLogsMetadata, SpendLogsPayload
from litellm.proxy._types import SpendLogsMetadata, SpendLogsPayload, SpendLogsRouterMetadata
from litellm.proxy.spend_tracking.spend_log_error_logger import spend_log_error
from litellm.proxy.utils import PrismaClient, hash_token
from litellm.types.utils import (
@ -93,6 +93,24 @@ def _redact_logged_api_key(value: str | None, *, already_redacted: bool = False)
return hash_token(stripped)
def _get_router_metadata_for_spend_log(
metadata: Mapping[str, object] | None,
requested_model: str | None,
selected_model: str | None,
selected_provider: str | None,
router_correlation_id: str | None,
) -> SpendLogsRouterMetadata | None:
model_info: Final = metadata.get("model_info") if metadata is not None else None
if not isinstance(model_info, Mapping) or model_info.get("internal_router_model") is not True:
return None
return SpendLogsRouterMetadata(
requested_model=requested_model or None,
selected_model=selected_model or None,
selected_provider=selected_provider or None,
router_correlation_id=router_correlation_id,
)
def _get_spend_logs_metadata(
metadata: dict | None,
applied_guardrails: list[str] | None = None,
@ -109,6 +127,7 @@ def _get_spend_logs_metadata(
cost_breakdown: CostBreakdown | None = None,
litellm_call_id: str | None = None,
autorouter_savings: float | None = None,
router_metadata: SpendLogsRouterMetadata | None = None,
) -> SpendLogsMetadata:
if metadata is None:
return SpendLogsMetadata(
@ -148,13 +167,17 @@ def _get_spend_logs_metadata(
autorouter_savings=autorouter_savings,
litellm_gateway_injected_cache=None,
litellm_call_id=litellm_call_id,
router_metadata=router_metadata,
)
verbose_proxy_logger.debug(
"getting payload for SpendLogs, available keys in metadata: " + str(list(metadata.keys()))
)
# Filter the metadata dictionary to include only the specified keys
clean_metadata: Final = SpendLogsMetadata(**{key: metadata.get(key) for key in SpendLogsMetadata.__annotations__})
clean_metadata: Final = SpendLogsMetadata(
**{key: metadata.get(key) for key in SpendLogsMetadata.__annotations__ if key != "router_metadata"},
router_metadata=router_metadata,
)
_raw_key: Final = clean_metadata.get("user_api_key")
_trusted_hash: Final = metadata.get("user_api_key_hash")
_already_redacted: Final = (
@ -375,6 +398,20 @@ def get_logging_payload(kwargs, response_obj, start_time, end_time) -> SpendLogs
hidden_params: Final = standard_logging_payload.get("hidden_params", {})
litellm_overhead_time_ms = hidden_params.get("litellm_overhead_time_ms")
custom_llm_provider: Final = (
kwargs.get("custom_llm_provider")
or _sl_attribution_fallback(standard_logging_payload, "custom_llm_provider")
or None
)
raw_model: Final = cast(str, kwargs.get("model") or "")
model_name: Final = (
standard_logging_payload.get("model") if standard_logging_payload is not None else None
) or reconstruct_model_name(raw_model, custom_llm_provider, metadata or {})
litellm_call_id: Final = cast(
str | None,
kwargs.get("litellm_call_id") or litellm_params.get("litellm_call_id"),
)
# clean up litellm metadata
clean_metadata = _get_spend_logs_metadata(
metadata,
@ -433,9 +470,13 @@ def get_logging_payload(kwargs, response_obj, start_time, end_time) -> SpendLogs
autorouter_savings=(
standard_logging_payload.get("autorouter_savings", None) if standard_logging_payload is not None else None
),
litellm_call_id=cast(
str | None,
kwargs.get("litellm_call_id") or litellm_params.get("litellm_call_id"),
litellm_call_id=litellm_call_id,
router_metadata=_get_router_metadata_for_spend_log(
metadata=metadata,
requested_model=_model_group,
selected_model=model_name,
selected_provider=custom_llm_provider,
router_correlation_id=litellm_call_id,
),
)
@ -480,15 +521,6 @@ def get_logging_payload(kwargs, response_obj, start_time, end_time) -> SpendLogs
# Extract agent_id for A2A requests (set directly on model_call_details)
agent_id: Final[str | None] = kwargs.get("agent_id") or metadata.get("agent_id")
custom_llm_provider: Final = (
kwargs.get("custom_llm_provider")
or _sl_attribution_fallback(standard_logging_payload, "custom_llm_provider")
or None
)
raw_model: Final = cast(str, kwargs.get("model") or "")
model_name: Final = (
standard_logging_payload.get("model") if standard_logging_payload is not None else None
) or reconstruct_model_name(raw_model, custom_llm_provider, metadata or {})
try:
payload: Final[SpendLogsPayload] = SpendLogsPayload(

6
litellm/proxy/utils.py
View file

@ -645,7 +645,7 @@ class ProxyLogging:
self.max_parallel_request_limiter = _PROXY_MaxParallelRequestsHandler(self.internal_usage_cache)
self.max_budget_limiter = _PROXY_MaxBudgetLimiter()
self.cache_control_check = _PROXY_CacheControlCheck()
self.alerting: list | None = None
self.alerting: list[str] | None = None
self.alerting_threshold: float = 300 # default to 5 min. threshold
self.alert_types: list[AlertType] = DEFAULT_ALERT_TYPES
self.alert_to_webhook_url: dict | None = None
@ -2364,7 +2364,9 @@ class ProxyLogging:
# do nothing if alerting is not switched on (unless it's a soft_budget alert with team-specific emails)
return
if self.alerting is not None and ("slack" in self.alerting or "ms_teams" in self.alerting):
if self.alerting is not None and (
"slack" in self.alerting or "ms_teams" in self.alerting or "webhook" in self.alerting
):
if self.slack_alerting_instance is not None:
await self.slack_alerting_instance.budget_alerts(
type=type,

2
litellm/responses/litellm_completion_transformation/transformation.py
View file

@ -1629,6 +1629,8 @@ class LiteLLMCompletionResponsesConfig:
file_dict["file_id"] = file_id
if item.get("file_data"):
file_dict["file_data"] = item["file_data"]
if item.get("filename"):
file_dict["filename"] = item["filename"]
new_item: Final[dict[str, object]] = {"type": "file", "file": file_dict}
if "cache_control" in item:

226
litellm/responses/streaming_iterator.py
View file

@ -5,11 +5,11 @@ import json
import time
import traceback
import uuid
from collections.abc import Awaitable, Callable, Mapping, Sequence
from collections.abc import Awaitable, Callable, Iterable, Mapping, Sequence
from datetime import datetime
from functools import lru_cache
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, runtime_checkable
from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, overload, runtime_checkable
import httpx
from openai._streaming import SSEDecoder
@ -42,27 +42,14 @@ from litellm.types.utils import CallTypes
from litellm.utils import async_post_call_success_deployment_hook
if TYPE_CHECKING:
from litellm.caching.caching_handler import LLMCachingHandler
from litellm.proxy._types import UserAPIKeyAuth
from litellm.types.responses.streaming_websocket import (
PresidioGuardrailCallback,
ResponsesBackendWebSocket,
ResponsesClientWebSocket,
)
class _StreamCachingHandler(Protocol):
"""The ``_llm_caching_handler`` attached to a logging object, as this module uses it."""
original_function: Callable[..., object]
def _should_store_result_in_cache(
self, original_function: Callable[..., object], kwargs: Mapping[str, object]
) -> bool: ...
class PiiUnmaskingGuardrailCallback(PresidioGuardrailCallback, Protocol):
"""Guardrail callback that can also reverse its own masking, selected by
``llm_http_handler`` on exactly this attribute."""
def _unmask_pii_text(self, text: str, pii_tokens: Mapping[str, str]) -> str: ...
from litellm.types.router import LiteLLM_Params
class ProjectQuotaCallback(Protocol):
@ -94,6 +81,60 @@ def _is_str_mapping(value: object) -> TypeIs[dict[str, str]]: # guard-ok: verif
return _is_json_object(value) and all(isinstance(item, str) for item in value.values())
class _MutableJsonObject(Protocol):
@overload
def get(self, key: str, /) -> object | None: ...
@overload
def get(self, key: str, default: object, /) -> object: ...
def __getitem__(self, key: str, /) -> object: ...
def __setitem__(self, key: str, value: object, /) -> None: ...
def __contains__(self, key: object, /) -> bool: ...
def items(self) -> Iterable[tuple[str, object]]: ...
class _GetsLitellmParams(Protocol):
def __call__(self, key: str, default: Mapping[str, object], /) -> LiteLLM_Params: ...
class _PopsOptionalStr(Protocol):
def __call__(self, key: str, default: None, /) -> str | None: ...
class _UnmasksPiiText(Protocol):
def __call__(self, text: str, pii_tokens: Mapping[str, str]) -> str: ...
class _ShouldStoreResultInCache(Protocol):
def __call__(self, *, original_function: Callable[..., object] | None, kwargs: Mapping[str, object]) -> bool: ...
class _PostStreamingDeploymentHook(Protocol):
def __call__(
self,
*,
request_data: Mapping[str, object],
response_chunk: ResponsesAPIStreamingResponse,
call_type: CallTypes | None,
) -> Awaitable[ResponsesAPIStreamingResponse | None]: ...
@runtime_checkable
class _HasPostStreamingDeploymentHook(Protocol):
async_post_call_streaming_deployment_hook: _PostStreamingDeploymentHook
def _typed_gets_litellm_params(fn: _GetsLitellmParams) -> _GetsLitellmParams:
return fn
def _typed_pops_optional_str(fn: _PopsOptionalStr) -> _PopsOptionalStr:
return fn
_SHOULD_STORE_RESULT_IN_CACHE_ATTR: Final = "_should_store_result_in_cache"
_UNMASK_PII_TEXT_ATTR: Final = "_unmask_pii_text"
def _load_json_object(payload: str | bytes) -> dict[str, object]:
"""Parse a JSON payload that the caller consumes as an object."""
return json.loads(payload)
@ -220,7 +261,7 @@ class BaseResponsesAPIStreamingIterator:
# This matches the stream wrapper in litellm/litellm_core_utils/streaming_handler.py
_api_base: Final = get_api_base(
model=model or "",
optional_params=self.logging_obj.model_call_details.get("litellm_params", {}),
optional_params=_typed_gets_litellm_params(self.logging_obj.model_call_details.get)("litellm_params", {}),
)
self._hidden_params: dict[str, object] = {
"model_id": _model_id_from_metadata(litellm_metadata),
@ -422,15 +463,20 @@ class BaseResponsesAPIStreamingIterator:
end_time: Final = datetime.now()
if is_async:
asyncio.create_task(
self.logging_obj.dispatch_success_handlers(
logging_response,
start_time=self.start_time,
end_time=end_time,
cache_hit=self._completed_response_cache_hit,
prefer_async_handlers=True,
)
logging_coroutine: Final = self.logging_obj.dispatch_success_handlers(
logging_response,
start_time=self.start_time,
end_time=end_time,
cache_hit=self._completed_response_cache_hit,
prefer_async_handlers=True,
)
deferred_dispatch_armed: Final = getattr(self.logging_obj, "_on_deferred_stream_complete", None) is not None
if deferred_dispatch_armed:
# End-of-stream guardrail scans write guardrail_information after
# the terminal event; dispatching now would snapshot metadata early.
self.logging_obj._deferred_stream_complete_args = (logging_coroutine,)
else:
asyncio.create_task(logging_coroutine)
else:
run_async_function(
async_function=self.logging_obj.async_success_handler,
@ -549,7 +595,7 @@ class BaseResponsesAPIStreamingIterator:
if response_obj is None:
return
caching_handler: Final[_StreamCachingHandler | None] = getattr(self.logging_obj, "_llm_caching_handler", None)
caching_handler: Final[LLMCachingHandler | None] = getattr(self.logging_obj, "_llm_caching_handler", None)
if caching_handler is None:
return
@ -567,8 +613,11 @@ class BaseResponsesAPIStreamingIterator:
if preset_cache_key is not None:
request_kwargs["cache_key"] = preset_cache_key
if not caching_handler._should_store_result_in_cache( # pyright: ignore[reportPrivateUsage] # no public API
original_function=caching_handler.original_function,
should_store_result_in_cache: Final[_ShouldStoreResultInCache] = getattr(
caching_handler, _SHOULD_STORE_RESULT_IN_CACHE_ATTR
)
if not should_store_result_in_cache(
original_function=getattr(caching_handler, "original_function", None),
kwargs=request_kwargs,
):
return
@ -624,12 +673,15 @@ class BaseResponsesAPIStreamingIterator:
typed_call_type = None
request_data: Final = self.request_data or getattr(self.logging_obj, "model_call_details", {})
callbacks: Final = getattr(litellm, "callbacks", None) or []
callbacks: Final[Sequence[object]] = getattr(litellm, "callbacks", None) or []
hooks_ran = False
for callback in callbacks:
if hasattr(callback, "async_post_call_streaming_deployment_hook"):
if isinstance(callback, _HasPostStreamingDeploymentHook):
hooks_ran = True
result = await callback.async_post_call_streaming_deployment_hook(
post_streaming_hook: _PostStreamingDeploymentHook = (
callback.async_post_call_streaming_deployment_hook
)
result = await post_streaming_hook(
request_data=request_data,
response_chunk=chunk,
call_type=typed_call_type,
@ -1083,7 +1135,7 @@ class _HasModelDumpJson(Protocol):
def model_dump_json(self, *, exclude_none: bool = ...) -> str: ...
def _dump_response_object(obj: object) -> dict[str, Any]:
def _dump_response_object(obj: object) -> Mapping[str, object]:
if isinstance(obj, _HasModelDump):
return obj.model_dump()
if _is_json_object(obj):
@ -1113,21 +1165,20 @@ def _build_content_part_done_event(
item_id: str,
output_index: int,
content_index: int,
part_payload: dict[str, Any],
part_payload: Mapping[str, object],
) -> ResponsesAPIStreamingResponse | None:
openai_types: Final = _get_openai_response_types()
part_type: Final = part_payload.get("type")
part: PART_UNION_TYPES
if part_type == "output_text":
annotations: Final = [
openai_types.BaseLiteLLMOpenAIResponseObject(**annotation)
for annotation in part_payload.get("annotations", []) or []
]
part = openai_types.ContentPartDonePartOutputText(
type="output_text",
text=str(part_payload.get("text") or ""),
annotations=annotations,
logprobs=part_payload.get("logprobs"),
raw_annotations: Final[object] = part_payload.get("annotations", []) or []
part = openai_types.ContentPartDonePartOutputText.model_validate(
{
"type": "output_text",
"text": str(part_payload.get("text") or ""),
"annotations": raw_annotations,
"logprobs": part_payload.get("logprobs"),
}
)
elif part_type == "refusal":
part = openai_types.ContentPartDonePartRefusal(
@ -1157,7 +1208,7 @@ def _add_text_like_part_events(
item_id: str,
output_index: int,
content_index: int,
part_payload: dict[str, Any],
part_payload: Mapping[str, object],
chunk_size: int,
) -> None:
openai_types: Final = _get_openai_response_types()
@ -1174,16 +1225,19 @@ def _add_text_like_part_events(
delta=text[i : i + chunk_size],
)
)
annotations_payload: Final[Sequence[dict[str, object]]] = part_payload.get("annotations", []) or []
for annotation_index, annotation in enumerate(annotations_payload):
raw_annotation_items: Final = part_payload.get("annotations")
annotation_items: Final[Sequence[object]] = raw_annotation_items if _is_json_array(raw_annotation_items) else []
for annotation_index, annotation in enumerate(annotation_items):
events.append(
openai_types.OutputTextAnnotationAddedEvent(
type=openai_types.ResponsesAPIStreamEvents.OUTPUT_TEXT_ANNOTATION_ADDED,
item_id=item_id,
output_index=output_index,
content_index=content_index,
annotation_index=annotation_index,
annotation=annotation,
openai_types.OutputTextAnnotationAddedEvent.model_validate(
{
"type": openai_types.ResponsesAPIStreamEvents.OUTPUT_TEXT_ANNOTATION_ADDED,
"item_id": item_id,
"output_index": output_index,
"content_index": content_index,
"annotation_index": annotation_index,
"annotation": annotation,
}
)
)
events.append(
@ -1256,7 +1310,8 @@ def _build_synthetic_response_events(
)
if item_type == "message":
content_parts: Sequence[object] = output_item_payload.get("content", []) or []
raw_content_parts = output_item_payload.get("content")
content_parts: Sequence[object] = raw_content_parts if _is_json_array(raw_content_parts) else []
for content_index, part in enumerate(content_parts):
part_payload = _dump_response_object(part)
events.append(
@ -1304,8 +1359,9 @@ def _build_synthetic_response_events(
)
)
elif item_type == "reasoning":
summaries: Sequence[object] = output_item_payload.get("summary", []) or []
for summary_index, summary in enumerate(summaries):
raw_summary_items = output_item_payload.get("summary")
summary_items: Sequence[object] = raw_summary_items if _is_json_array(raw_summary_items) else []
for summary_index, summary in enumerate(summary_items):
summary_payload = _dump_response_object(summary)
summary_text = str(summary_payload.get("text") or "")
for i in range(0, len(summary_text), chunk_size):
@ -1476,7 +1532,7 @@ class ResponsesWebSocketStreaming:
user_api_key_dict: UserAPIKeyAuth | None = None,
request_data: dict[str, object] | None = None,
first_message: str | None = None,
guardrail_callbacks: list[PiiUnmaskingGuardrailCallback] | None = None,
guardrail_callbacks: Sequence[PresidioGuardrailCallback] | None = None,
output_guardrail_callbacks: list[PresidioGuardrailCallback] | None = None,
quota_callbacks: Sequence[ProjectQuotaCallback] | None = None,
authorized_model: str | None = None,
@ -1486,17 +1542,17 @@ class ResponsesWebSocketStreaming:
self.logging_obj = logging_obj
self.user_api_key_dict = user_api_key_dict
self.request_data: dict[str, object] = request_data or {}
self.messages: list[dict[str, object]] = []
self.messages: list[_MutableJsonObject] = []
self.input_messages: list[dict[str, object]] = []
self.first_message = first_message
self.guardrail_callbacks: list[PiiUnmaskingGuardrailCallback] = guardrail_callbacks or []
self.guardrail_callbacks: Sequence[PresidioGuardrailCallback] = guardrail_callbacks or []
self.output_guardrail_callbacks: list[PresidioGuardrailCallback] = output_guardrail_callbacks or []
self.quota_callbacks: tuple[ProjectQuotaCallback, ...] = tuple(quota_callbacks) if quota_callbacks else ()
# Model name authorized at connection time; enforced on every
# response.create frame to prevent deployment-substitution attacks.
self.authorized_model: str | None = authorized_model
def _should_store_event(self, event_obj: Mapping[str, object]) -> bool:
def _should_store_event(self, event_obj: _MutableJsonObject) -> bool:
return event_obj.get("type") in RESPONSES_WS_LOGGED_EVENT_TYPES
def _store_event(self, event: str | bytes | dict[str, object]) -> None:
@ -1610,7 +1666,7 @@ class ResponsesWebSocketStreaming:
finally:
await self._log_messages()
def _enforce_authorized_model(self, msg_obj: dict[str, object]) -> bool:
def _enforce_authorized_model(self, msg_obj: _MutableJsonObject) -> bool:
"""
Overwrite any ``model`` field in a ``response.create`` frame with the
connection-authorized model to prevent deployment-substitution attacks.
@ -1679,7 +1735,7 @@ class ResponsesWebSocketStreaming:
# forwarded unmasked regardless of where the client places it.
nested_candidate = msg_obj.get("response")
nested_response = nested_candidate if _is_json_object(nested_candidate) else None
text_containers: list[tuple[dict[str, object], str]] = []
text_containers: list[tuple[_MutableJsonObject, str]] = []
for container in (msg_obj, nested_response):
if container is None:
continue
@ -1786,6 +1842,7 @@ class ResponsesWebSocketStreaming:
return response_str
cb: Final = self.guardrail_callbacks[0]
unmask_pii_text: Final[_UnmasksPiiText] = getattr(cb, _UNMASK_PII_TEXT_ATTR)
event_type: Final = evt_obj.get("type")
if event_type == "response.completed":
@ -1805,9 +1862,7 @@ class ResponsesWebSocketStreaming:
continue
text = content_block.get("text")
if isinstance(text, str):
unmasked = cb._unmask_pii_text( # pyright: ignore[reportPrivateUsage] # no public unmasker
text, pii_tokens
)
unmasked = unmask_pii_text(text, pii_tokens)
if unmasked != text:
content_block["text"] = unmasked
modified = True
@ -1816,9 +1871,7 @@ class ResponsesWebSocketStreaming:
if event_type in self._DELTA_EVENT_TYPES:
delta: Final = evt_obj.get("delta")
if isinstance(delta, str):
unmasked = cb._unmask_pii_text( # pyright: ignore[reportPrivateUsage] # no public unmasker
delta, pii_tokens
)
unmasked = unmask_pii_text(delta, pii_tokens)
if unmasked != delta:
evt_obj["delta"] = unmasked
return json.dumps(evt_obj)
@ -2020,7 +2073,7 @@ class ManagedResponsesWebSocketHandler:
model: str,
logging_obj: LiteLLMLoggingObj,
user_api_key_dict: UserAPIKeyAuth | None = None,
litellm_metadata: dict[str, Any] | None = None,
litellm_metadata: Mapping[str, object] | None = None,
api_key: str | None = None,
api_base: str | None = None,
timeout: float | None = None,
@ -2033,10 +2086,11 @@ class ManagedResponsesWebSocketHandler:
self.model = model
self.logging_obj = logging_obj
self.user_api_key_dict = user_api_key_dict
self.litellm_metadata: dict[str, Any] = litellm_metadata or {}
self.model_group: str | None = self.litellm_metadata.get("model_group") or self.litellm_metadata.get(
self.litellm_metadata: Mapping[str, object] = litellm_metadata or {}
raw_model_group: Final = self.litellm_metadata.get("model_group") or self.litellm_metadata.get(
"deployment_model_name"
)
self.model_group: str | None = raw_model_group if isinstance(raw_model_group, str) else None
self.api_key = api_key
self.api_base = api_base
self.timeout = timeout
@ -2057,7 +2111,7 @@ class ManagedResponsesWebSocketHandler:
# ------------------------------------------------------------------
@staticmethod
def _serialize_chunk(chunk: Any) -> str | None:
def _serialize_chunk(chunk: object) -> str | None:
"""Serialize a streaming chunk to a JSON string for WebSocket transmission."""
try:
if isinstance(chunk, _HasModelDumpJson):
@ -2100,7 +2154,7 @@ class ManagedResponsesWebSocketHandler:
self._session_history[response_id] = messages
@staticmethod
def _extract_response_id(completed_event: dict[str, object]) -> str | None:
def _extract_response_id(completed_event: _MutableJsonObject) -> str | None:
"""
Pull the raw (decoded) response ID out of a ``response.completed`` event.
Returns *None* if the event doesn't contain a usable ID.
@ -2115,7 +2169,7 @@ class ManagedResponsesWebSocketHandler:
@staticmethod
def _extract_output_messages(
completed_event: dict[str, object],
completed_event: _MutableJsonObject,
) -> list[dict[str, object]]:
"""
Convert the output items in a ``response.completed`` event into
@ -2172,7 +2226,7 @@ class ManagedResponsesWebSocketHandler:
# _process_response_create sub-methods
# ------------------------------------------------------------------
async def _parse_message(self, raw_message: str) -> dict[str, object] | None:
async def _parse_message(self, raw_message: str) -> _MutableJsonObject | None:
"""Parse raw WS text; return the message dict or None (JSON error / ignored type)."""
try:
msg_obj: Final = _load_json_object(raw_message)
@ -2185,7 +2239,7 @@ class ManagedResponsesWebSocketHandler:
return msg_obj
@staticmethod
def _is_warmup_frame(msg_obj: dict[str, object]) -> bool:
def _is_warmup_frame(msg_obj: _MutableJsonObject) -> bool:
"""Return True for a response.create whose generate flag is false."""
nested: Final = msg_obj.get("response")
source: Final = nested if _is_json_object(nested) and nested else msg_obj
@ -2201,13 +2255,13 @@ class ManagedResponsesWebSocketHandler:
return str(raw_id).startswith(_WARMUP_RESPONSE_ID_PREFIX)
@staticmethod
def _warmup_source_params(msg_obj: dict[str, object]) -> dict[str, object]:
def _warmup_source_params(msg_obj: _MutableJsonObject) -> dict[str, object]:
nested: Final = msg_obj.get("response")
if _is_json_object(nested) and nested:
return nested
return {k: v for k, v in msg_obj.items() if k != "type"}
def _build_warmup_response(self, msg_obj: dict[str, object]) -> dict[str, object]:
def _build_warmup_response(self, msg_obj: _MutableJsonObject) -> dict[str, object]:
"""Build a minimal completed Responses API object for a warmup ack."""
source: Final = self._warmup_source_params(msg_obj)
wire_model: Final = source.get("model") or self.model_group or self.model
@ -2225,7 +2279,7 @@ class ManagedResponsesWebSocketHandler:
},
}
async def _send_warmup_ack(self, msg_obj: dict[str, object]) -> None:
async def _send_warmup_ack(self, msg_obj: _MutableJsonObject) -> None:
"""
Acknowledge a generate=false prewarm without calling the provider.
@ -2248,7 +2302,7 @@ class ManagedResponsesWebSocketHandler:
await self.websocket.send_text(serialized)
@staticmethod
def _build_base_call_kwargs(msg_obj: dict[str, object]) -> dict[str, Any]:
def _build_base_call_kwargs(msg_obj: _MutableJsonObject) -> dict[str, Any]:
"""
Extract Responses API params from the event, handling both wire formats:
Nested: {"type": "response.create", "response": {"input": [...], ...}}
@ -2357,7 +2411,7 @@ class ManagedResponsesWebSocketHandler:
call_kwargs.setdefault("litellm_params", {})
call_kwargs["litellm_params"]["proxy_server_request"] = proxy_server_request
async def _stream_and_forward(self, model: str, call_kwargs: dict[str, Any]) -> dict[str, object] | None:
async def _stream_and_forward(self, model: str, call_kwargs: dict[str, Any]) -> _MutableJsonObject | None:
"""
Stream ``litellm.aresponses`` and forward every chunk over the WebSocket.
@ -2365,7 +2419,7 @@ class ManagedResponsesWebSocketHandler:
directly (before serialization) to avoid a redundant JSON round-trip on
every chunk. Returns the completed event dict, or ``None``.
"""
completed_event: dict[str, object] | None = (
completed_event: _MutableJsonObject | None = (
None # rebind-ok: captures the completed event once the stream yields it
)
stream_response: Final = await litellm.aresponses(model=model, **call_kwargs)
@ -2391,7 +2445,7 @@ class ManagedResponsesWebSocketHandler:
def _save_turn_history(
self,
completed_event: dict[str, object] | None,
completed_event: _MutableJsonObject | None,
prior_history: list[dict[str, object]],
current_messages: list[dict[str, object]],
) -> None:
@ -2464,12 +2518,14 @@ class ManagedResponsesWebSocketHandler:
# reuse the router-resolved self.model; passing the alias raw to
# litellm.aresponses fails in get_llm_provider. A genuinely different
# provider-prefixed per-frame model is still honored.
requested_model: Final[str | None] = call_kwargs.pop("model", None)
requested_model: Final[str | None] = _typed_pops_optional_str(call_kwargs.pop)("model", None)
model: Final[str] = (
self.model if requested_model is None or requested_model == self.model_group else requested_model
)
previous_response_id: Final[str | None] = call_kwargs.pop("previous_response_id", None)
previous_response_id: Final[str | None] = _typed_pops_optional_str(call_kwargs.pop)(
"previous_response_id", None
)
current_messages: Final = self._input_to_messages(call_kwargs.get("input"))
# Fetch history once; reused in both _apply_history and _save_turn_history

212
litellm/router_strategy/tag_based_routing.py
View file

@ -8,16 +8,14 @@ Use this to route requests between Teams
"""
import re
from collections.abc import Iterable, Mapping, Sequence
from collections.abc import Mapping, Sequence
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, TypedDict
from typing_extensions import ReadOnly
from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, overload
from litellm._logging import verbose_logger
from litellm.constants import CONSUMED_REQUEST_TAGS_METADATA_KEY
from litellm.litellm_core_utils.core_helpers import get_metadata_variable_name_from_kwargs
from litellm.types.router import ConsumedRequestTagsStamp, RouterErrors
from litellm.types.router import ConsumedRequestTagsStamp, DeploymentTypedDict, RouterErrors
if TYPE_CHECKING:
from litellm.router import Router as _Router
@ -27,34 +25,63 @@ else:
LitellmRouter = Any
class _TagRoutingLitellmParams(TypedDict, total=False):
tags: ReadOnly[Sequence[str] | None]
tag_regex: ReadOnly[Sequence[str] | None]
class _TagLitellmParamsLike(Protocol):
@overload
def get(self, key: Literal["tags"], /) -> Sequence[str] | None: ...
@overload
def get(self, key: Literal["tags"], default: Sequence[str], /) -> Sequence[str]: ...
@overload
def get(self, key: Literal["tag_regex"], /) -> Sequence[str] | None: ...
class _TagRoutingDeployment(TypedDict, total=False):
model_name: ReadOnly[str]
litellm_params: ReadOnly[_TagRoutingLitellmParams]
model_info: ReadOnly[Mapping[str, object] | None]
class _ModelInfoLike(Protocol):
@overload
def get(self, key: Literal["allow_fail_open"], /) -> bool | None: ...
@overload
def get(self, key: Literal["enable_tag_filtering"], /) -> bool | None: ...
class _TagRoutingMatchStamp(TypedDict):
matched_deployment: ReadOnly[str | None]
matched_via: ReadOnly[str]
matched_value: ReadOnly[str]
request_tags: ReadOnly[Sequence[str]]
user_agent: ReadOnly[str]
class _DeploymentLike(Protocol):
@overload
def get(self, key: Literal["litellm_params"], default: Mapping[str, object], /) -> _TagLitellmParamsLike: ...
@overload
def get(self, key: Literal["model_info"], /) -> _ModelInfoLike | None: ...
@overload
def get(self, key: Literal["model_name"], /) -> object: ...
class _TagRoutingMetadata(TypedDict, total=False):
tags: ReadOnly[Sequence[str] | None]
inherited_tags: ReadOnly[Sequence[str] | None]
user_agent: ReadOnly[str]
tag_routing: ReadOnly[_TagRoutingMatchStamp]
_consumed_request_tags: ReadOnly[object]
class _MetadataLike(Protocol):
@overload
def get(self, key: Literal["tags"], /) -> Sequence[str] | None: ...
@overload
def get(self, key: Literal["tags"], default: Sequence[str], /) -> Sequence[str] | None: ...
@overload
def get(self, key: Literal["user_agent"], default: str, /) -> str: ...
@overload
def get(self, key: Literal["inherited_tags"], /) -> object: ...
def __contains__(self, key: object, /) -> bool: ...
def __setitem__(self, key: Literal["tag_routing"], value: Mapping[str, object], /) -> None: ...
_EMPTY_MODEL_INFO: Final[Mapping[str, object]] = MappingProxyType({})
class _NestedLitellmParamsLike(Protocol):
def get(
self, key: Literal["metadata", "litellm_metadata"], default: Mapping[str, object], /
) -> _MetadataLike | None: ...
class _RequestKwargsLike(Protocol):
@overload
def get(self, key: Literal["enable_tag_filtering"], /) -> bool | None: ...
@overload
def get(self, key: Literal["metadata", "litellm_metadata"], /) -> _MetadataLike | None: ...
def __contains__(self, key: object, /) -> bool: ...
@overload
def __getitem__(self, key: Literal["metadata", "litellm_metadata"], /) -> _MetadataLike: ...
@overload
def __getitem__(self, key: Literal["litellm_params"], /) -> _NestedLitellmParamsLike: ...
_DeploymentPool = Sequence[_DeploymentLike] | Mapping[_DeploymentLike, object]
def _is_valid_deployment_tag_regex(
@ -109,11 +136,11 @@ def is_valid_deployment_tag(
def _match_deployment(
deployment: _TagRoutingDeployment,
request_tags: Sequence[str] | None,
header_strings: Sequence[str],
deployment: _DeploymentLike,
request_tags: list[str] | None,
header_strings: list[str],
match_any: bool,
) -> Mapping[str, str] | None:
) -> dict[str, str] | None:
"""
Determine whether *deployment* matches the current request.
@ -198,38 +225,38 @@ def _split_tags(tags: Sequence[str]) -> tuple[tuple[str, ...], list[str], tuple[
def _exclude_deployments(
deployments: Iterable[_TagRoutingDeployment],
deployments: _DeploymentPool,
excluded_set: frozenset[str],
) -> list[_TagRoutingDeployment]:
) -> Sequence[_DeploymentLike]:
if not excluded_set:
return list(deployments)
return [d for d in deployments if not excluded_set.intersection(d.get("litellm_params", {}).get("tags") or [])]
def _require_all_tags(
deployments: Iterable[_TagRoutingDeployment],
deployments: _DeploymentPool,
required_set: frozenset[str],
) -> tuple[_TagRoutingDeployment, ...]:
) -> tuple[_DeploymentLike, ...]:
if not required_set:
return tuple(deployments)
return tuple(d for d in deployments if required_set.issubset(d.get("litellm_params", {}).get("tags") or []))
def _default_tagged_pool(
deployments: Iterable[_TagRoutingDeployment],
) -> tuple[_TagRoutingDeployment, ...]:
deployments: _DeploymentPool,
) -> tuple[_DeploymentLike, ...]:
defaults: Final = tuple(d for d in deployments if "default" in (d.get("litellm_params", {}).get("tags") or []))
return defaults if defaults else tuple(deployments)
def _known_tag_values(deployments: Iterable[_TagRoutingDeployment]) -> frozenset[str]:
def _known_tag_values(deployments: _DeploymentPool) -> frozenset[str]:
return frozenset(
tag for d in deployments for tag in (d.get("litellm_params", _TagRoutingLitellmParams()).get("tags") or ())
tag for d in deployments for tag in (d.get("litellm_params", MappingProxyType({})).get("tags") or ())
)
def _unknown_required_tag_hides_an_answer(
healthy_deployments: Iterable[_TagRoutingDeployment],
healthy_deployments: _DeploymentPool,
excluded_set: frozenset[str],
required_set: frozenset[str],
routing_confirmed: frozenset[str],
@ -253,23 +280,23 @@ def _unknown_required_tag_hides_an_answer(
def _chain_allows_fail_open(
healthy_deployments: Iterable[_TagRoutingDeployment],
healthy_deployments: _DeploymentPool,
excluded_set: frozenset[str],
required_set: frozenset[str],
routing_confirmed: frozenset[str],
) -> bool:
if _unknown_required_tag_hides_an_answer(healthy_deployments, excluded_set, required_set, routing_confirmed):
return False
return any((d.get("model_info") or _EMPTY_MODEL_INFO).get("allow_fail_open") is True for d in healthy_deployments)
return any((d.get("model_info") or {}).get("allow_fail_open") is True for d in healthy_deployments)
def _trusted_only_pool(
healthy_deployments: Iterable[_TagRoutingDeployment],
healthy_deployments: _DeploymentPool,
excluded_set: frozenset[str],
required_set: frozenset[str],
inherited_excluded_set: frozenset[str] | None,
inherited_required_set: frozenset[str] | None,
) -> tuple[_TagRoutingDeployment, ...]:
) -> tuple[_DeploymentLike, ...]:
# inherited_*_set is None only when this request carries no origin information
# at all (e.g. direct SDK Router usage, bypassing the proxy layer that
# populates metadata.inherited_tags) -- treat every constraint as
@ -296,8 +323,8 @@ def _trusted_only_pool(
def _resolve_or_fail_open(
pool: Sequence[_TagRoutingDeployment],
healthy_deployments: Iterable[_TagRoutingDeployment],
pool: Sequence[_DeploymentLike],
healthy_deployments: _DeploymentPool,
excluded_set: frozenset[str],
required_set: frozenset[str],
inherited_excluded_set: frozenset[str] | None,
@ -305,7 +332,7 @@ def _resolve_or_fail_open(
routing_confirmed: frozenset[str],
model: str,
request_tags: object,
) -> tuple[_TagRoutingDeployment, ...]:
) -> tuple[_DeploymentLike, ...]:
if pool:
return tuple(pool)
if _chain_allows_fail_open(healthy_deployments, excluded_set, required_set, routing_confirmed):
@ -325,7 +352,7 @@ def _resolve_or_fail_open(
def _resolve_constraint_only_pool(
healthy_deployments: Iterable[_TagRoutingDeployment],
healthy_deployments: _DeploymentPool,
excluded_set: frozenset[str],
required_set: frozenset[str],
inherited_excluded_set: frozenset[str] | None,
@ -333,7 +360,7 @@ def _resolve_constraint_only_pool(
routing_confirmed: frozenset[str],
model: str,
request_tags: object,
) -> tuple[_TagRoutingDeployment, ...]:
) -> tuple[_DeploymentLike, ...]:
pool: Final = (
_require_all_tags(_exclude_deployments(healthy_deployments, excluded_set), required_set)
if required_set
@ -355,8 +382,8 @@ def _resolve_constraint_only_pool(
def _all_deployments_or_fallback(
llm_router_instance: LitellmRouter,
model: str,
fallback: Iterable[_TagRoutingDeployment],
) -> Iterable[_TagRoutingDeployment]:
fallback: _DeploymentPool,
) -> Sequence[_DeploymentLike | DeploymentTypedDict] | Mapping[_DeploymentLike, object]:
try:
return llm_router_instance._get_all_deployments(model_name=model)
except Exception: # noqa: BLE001 # fail safe toward today's healthy-only behavior on lookup errors
@ -366,8 +393,8 @@ def _all_deployments_or_fallback(
def _chain_tag_filtering_override(
llm_router_instance: LitellmRouter,
model: str,
healthy_deployments: Iterable[_TagRoutingDeployment],
) -> object:
healthy_deployments: _DeploymentPool,
) -> bool | None:
# Resolved from every deployment configured for this model group, not just the
# ones that survived cooldown/health filtering (async_get_healthy_deployments
# filters cooldowns before calling get_deployments_for_tag) -- otherwise the
@ -379,14 +406,14 @@ def _chain_tag_filtering_override(
# than crashing the request.
all_deployments: Final = _all_deployments_or_fallback(llm_router_instance, model, healthy_deployments)
for d in all_deployments:
value = (d.get("model_info") or _EMPTY_MODEL_INFO).get("enable_tag_filtering")
value = (d.get("model_info") or MappingProxyType({})).get("enable_tag_filtering")
if value is not None:
return value
return None
def _inherited_constraint_sets(
inherited_tags: Sequence[str] | None, routing_prefix: str
inherited_tags: object, routing_prefix: str
) -> tuple[frozenset[str] | None, frozenset[str] | None]:
# None means no origin information is available at all (e.g. this request
# bypassed the proxy layer that populates metadata.inherited_tags, as direct
@ -417,43 +444,42 @@ def _tag_known_to_group(
if tag_set & routing_confirmed:
return True
try:
all_deployments: Final[Sequence[_TagRoutingDeployment]] = llm_router_instance._get_all_deployments(
model_name=model
)
all_deployments: Final = llm_router_instance._get_all_deployments(model_name=model)
except Exception: # noqa: BLE001 # fail safe toward "unrecognized" so lookup errors preserve the existing silent-fallback behavior
return False
return any(
tag_set.intersection(d.get("litellm_params", _TagRoutingLitellmParams()).get("tags") or ())
for d in all_deployments
tag_set.intersection(d.get("litellm_params", MappingProxyType({})).get("tags") or ()) for d in all_deployments
)
def _request_tags_after_router_consumption(metadata: _TagRoutingMetadata, model: str) -> Sequence[str] | None:
def _request_tags_after_router_consumption(metadata: object, model: str) -> Sequence[str] | None:
# The pre-routing hook stamps which tags selected the router it rewrote the request
# to: those tags already did their job and must not also constrain deployment choice
# inside the routed group. The request's other tags still apply there, on top of the
# inherited_tags snapshot that keeps key/team policy applying. Every other model
# group keeps the full list.
stamp: Final = metadata.get(CONSUMED_REQUEST_TAGS_METADATA_KEY)
if not isinstance(metadata, Mapping):
return None
typed_metadata: Final[Mapping[str, object]] = metadata
request_tags: Final = _tags_in_metadata(typed_metadata)
stamp: Final = typed_metadata.get(CONSUMED_REQUEST_TAGS_METADATA_KEY)
if not isinstance(stamp, ConsumedRequestTagsStamp) or stamp.model_group != model:
return metadata.get("tags")
request_tags: Final = metadata.get("tags")
leftover: Final = tuple(
tag for tag in (request_tags if isinstance(request_tags, (list, tuple)) else ()) if tag not in stamp.tags
)
inherited_tags: Final = metadata.get("inherited_tags")
return request_tags
leftover: Final = tuple(tag for tag in request_tags if tag not in stamp.tags)
inherited_tags: Final = typed_metadata.get("inherited_tags")
if not isinstance(inherited_tags, (list, tuple)):
return leftover or None
return tuple(dict.fromkeys((*leftover, *inherited_tags)))
typed_inherited_tags: Final[Sequence[object]] = inherited_tags
return tuple(dict.fromkeys((*leftover, *(tag for tag in typed_inherited_tags if isinstance(tag, str)))))
async def get_deployments_for_tag(
llm_router_instance: LitellmRouter,
model: str, # used to raise the correct error
healthy_deployments: list[Any] | dict[Any, Any],
request_kwargs: dict[Any, Any] | None = None,
healthy_deployments: _DeploymentPool,
request_kwargs: _RequestKwargsLike | None = None,
metadata_variable_name: Literal["metadata", "litellm_metadata"] = "metadata",
):
) -> _DeploymentPool:
"""
Returns a list of deployments that match the requested model and tags in the request.
@ -486,8 +512,7 @@ async def get_deployments_for_tag(
verbose_logger.debug("request metadata: %s", request_kwargs.get(metadata_variable_name))
if metadata_variable_name in request_kwargs:
metadata: Final[_TagRoutingMetadata] = request_kwargs[metadata_variable_name]
stampable_metadata: Final[dict[str, object]] = request_kwargs[metadata_variable_name]
metadata: Final = request_kwargs[metadata_variable_name]
request_tags: Final = _request_tags_after_router_consumption(metadata, model)
match_any: Final = llm_router_instance.tag_filtering_match_any
routing_prefix: Final = llm_router_instance.tag_routing_prefix or ""
@ -532,25 +557,25 @@ async def get_deployments_for_tag(
request_tags,
)
new_healthy_deployments: Final[list[_TagRoutingDeployment]] = []
default_deployments: Final[list[_TagRoutingDeployment]] = []
if has_positive_filter:
verbose_logger.debug(
"get_deployments_for_tag routing: request_tags=%s user_agent=%s",
request_tags,
user_agent,
)
for deployment in candidates:
deployment_tags = deployment.get("litellm_params", {}).get("tags")
match_result = _match_deployment(
deployment=deployment,
request_tags=positive_tags,
header_strings=header_strings,
match_any=match_any,
deployment_matches: Final = tuple(
(
deployment,
_match_deployment(
deployment=deployment,
request_tags=positive_tags,
header_strings=header_strings,
match_any=match_any,
),
)
for deployment in candidates
)
for deployment, match_result in deployment_matches:
if match_result is not None:
verbose_logger.debug(
"tag routing match: deployment=%s matched_via=%s matched_value=%s",
@ -559,17 +584,17 @@ async def get_deployments_for_tag(
match_result["matched_value"],
)
if "tag_routing" not in metadata:
stampable_metadata["tag_routing"] = {
metadata["tag_routing"] = {
"matched_deployment": deployment.get("model_name"),
"matched_via": match_result["matched_via"],
"matched_value": match_result["matched_value"],
"request_tags": request_tags or [],
"user_agent": user_agent,
}
new_healthy_deployments.append(deployment)
if deployment_tags and "default" in deployment_tags:
default_deployments.append(deployment)
new_healthy_deployments: Final = [d for d, result in deployment_matches if result is not None]
default_deployments: Final = [
d for d, _ in deployment_matches if "default" in (d.get("litellm_params", {}).get("tags") or ())
]
if len(new_healthy_deployments) == 0 and len(default_deployments) == 0:
return _resolve_or_fail_open(
@ -604,10 +629,11 @@ async def get_deployments_for_tag(
return new_healthy_deployments if len(new_healthy_deployments) > 0 else default_deployments
# for Untagged requests use default deployments if set
_default_deployments_with_tags: Final[list[_TagRoutingDeployment]] = []
for deployment in healthy_deployments:
if "default" in deployment.get("litellm_params", {}).get("tags", []):
_default_deployments_with_tags.append(deployment)
_default_deployments_with_tags: Final = [
deployment
for deployment in healthy_deployments
if "default" in deployment.get("litellm_params", {}).get("tags", [])
]
if len(_default_deployments_with_tags) > 0:
return _default_deployments_with_tags

36
litellm/types/guardrails.py
View file

@ -1,5 +1,7 @@
from collections.abc import Mapping
from datetime import datetime
from enum import Enum
from types import MappingProxyType
from typing import Any, Final, Literal
from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
@ -550,6 +552,40 @@ class BedrockGuardrailConfigModel(BaseModel):
)
class BedrockGuardrailStreamingParams(BaseModel):
streaming_buffer_until_moderated: bool = Field(
default=True,
description="If True (default), withhold every streamed chunk until the end-of-stream "
"ApplyGuardrail scan passes, so no flagged content reaches the client before a block. "
"If False, chunks stream through unbuffered, so flagged content can reach the client "
"before the scan finishes; a flagged scan still ends the stream, with a block message "
"when disable_exception_on_block is true and an in-stream error frame otherwise.",
)
streaming_sampling_rate: int = Field(
default=5,
ge=1,
description="When not buffering and not end-of-stream-only, scan the accumulated response "
"every Nth streamed chunk. Each sampled scan is a full ApplyGuardrail call that delays "
"that chunk, so lower values add latency and AWS text-unit cost.",
)
streaming_end_of_stream_only: bool = Field(
default=False,
description="When not buffering, skip per-chunk sampling and run one ApplyGuardrail scan "
"on the assembled response at end of stream. Combined with "
"streaming_buffer_until_moderated=false the full response streams live before the scan "
"and the scan result lands in guardrail_information; a flagged response still ends the "
"stream with a block message (disable_exception_on_block=true) or an error frame.",
)
@classmethod
def from_extras(cls, extras: Mapping[str, object] | None) -> "BedrockGuardrailStreamingParams":
if not extras:
return cls()
return cls.model_validate(
MappingProxyType({name: extras[name] for name in cls.model_fields if extras.get(name) is not None})
)
class LakeraV2GuardrailConfigModel(BaseModel):
"""Configuration parameters for the Lakera AI v2 guardrail"""

4
litellm/types/proxy/guardrails/guardrail_hooks/prompt_security.py
View file

@ -12,6 +12,10 @@ class PromptSecurityGuardrailConfigModel(GuardrailConfigModel):
default=None,
description="The API base for the Prompt Security guardrail. If not provided, the `PROMPT_SECURITY_API_BASE` environment variable is used.",
)
file_sanitization_fail_open: bool = Field(
default=True,
description="Whether file sanitization timeouts allow the original file through instead of blocking the request.",
)
@staticmethod
def ui_friendly_name() -> str:

5
litellm/types/router.py
View file

@ -189,6 +189,11 @@ class ModelInfo(MirroredPricingParams):
# router-wide default.
enable_tag_filtering: bool | None = None
# when True, calls routed to this deployment persist a router_metadata block
# (requested model group, selected model + provider, router correlation id)
# in the spend log row's metadata. Set it on every deployment of the group.
internal_router_model: bool | None = None
def __init__(self, id: str | int | None = None, **params) -> None:
if id is None:
id = str(uuid.uuid4()) # Generate a UUID if id is None or not provided

12
migrations/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:a31344ab2cb8618db84f535eec56f76f6178b142cb92cb2e48676cc2dcebea72
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin
@ -35,7 +35,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
RUN for i in 1 2 3; do \
apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done
@ -56,7 +56,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv sync --frozen --no-install-project --no-install-workspace --no-default-groups --no-editable \
--extra proxy \
--extra extra_proxy \
--python python3
--python python3.13
# Stage 2 — copy source and install the project + workspace members.
COPY . .
@ -65,7 +65,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv sync --frozen --no-default-groups --no-editable \
--extra proxy \
--extra extra_proxy \
--python python3
--python python3.13
COPY migrations/run.py /app/run.py
@ -87,7 +87,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
RUN for i in 1 2 3; do \
apk add --no-cache bash openssl tzdata python3 nodejs libsndfile libatomic && break; \
apk add --no-cache bash openssl tzdata python-3.13 nodejs libsndfile libatomic && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done

295
model_prices_and_context_window.json
View file

File diff suppressed because it is too large Load diff

14
ruff-strict-budget.json
View file

@ -1,6 +1,6 @@
{
"ANN001": {
"limit": 2998
"limit": 2995
},
"ANN002": {
"limit": 71
@ -9,7 +9,7 @@
"limit": 809
},
"ANN201": {
"limit": 2003
"limit": 2002
},
"ANN202": {
"limit": 845
@ -24,7 +24,7 @@
"limit": 133
},
"ANN401": {
"limit": 597
"limit": 587
},
"ASYNC230": {
"limit": 11
@ -123,7 +123,7 @@
"limit": 12
},
"PERF403": {
"limit": 34
"limit": 33
},
"PIE804": {
"limit": 18
@ -177,7 +177,7 @@
"limit": 8
},
"RUF019": {
"limit": 32
"limit": 31
},
"RUF046": {
"limit": 4
@ -198,7 +198,7 @@
"limit": 56
},
"SIM102": {
"limit": 315
"limit": 314
},
"SIM103": {
"limit": 119
@ -231,7 +231,7 @@
"limit": 5
},
"TID251": {
"limit": 1111
"limit": 1108
},
"TRY002": {
"limit": 524

348
scripts/auto-close-duplicates.test.ts Normal file
View file

@ -0,0 +1,348 @@
import { describe, expect, test } from "bun:test";
import {
CLOSED_MARKER,
REOPEN_COMMENT,
candidateNumbers,
duplicateTarget,
normalizeTitle,
pendingNotice,
readConfig,
reopenTarget,
sweepClosedIssue,
sweepIssue,
type Comment,
type GitHubApi,
type Issue,
type Reaction,
type SweepConfig,
} from "./auto-close-duplicates";
const NOW = new Date("2026-09-04T09:00:00Z");
const DAY_MS = 24 * 60 * 60 * 1000;
const daysAgo = (days: number): string => new Date(NOW.getTime() - days * DAY_MS).toISOString();
const issue = (number: number, title: string, overrides: Partial<Issue> = {}): Issue => ({
number,
title,
state: "open",
user: { login: "reporter" },
...overrides,
});
const notice = (candidates: readonly number[], createdAt: string, overrides: Partial<Comment> = {}): Comment => ({
id: 900,
body: `<!-- litellm:potential-duplicate candidates=${candidates.join(",")}, -->\n**Potential duplicate detected**`,
created_at: createdAt,
user: { type: "Bot", login: "github-actions[bot]" },
...overrides,
});
const humanComment = (createdAt: string, body = "It is not the same thing", login = "reporter"): Comment => ({
id: 901,
body,
created_at: createdAt,
user: { type: "User", login },
});
const config: SweepConfig = { repo: "BerriAI/litellm", graceDays: 3, dryRun: false, now: NOW };
describe("normalizeTitle", () => {
test("drops the template prefix, case, and punctuation", () => {
expect(normalizeTitle("[Bug]: Gemma 4-e4b fails on Vertex!")).toBe("gemma 4 e4b fails on vertex");
expect(normalizeTitle("[Feature]: ")).toBe("");
});
});
describe("candidateNumbers", () => {
test("reads only the marker field, keeps older issues, sorted ascending and deduplicated", () => {
const body = "<!-- litellm:potential-duplicate candidates=40,10,30,10, -->\n- #1 - see #1 (100% similar)";
expect(candidateNumbers(body, 35)).toEqual([10, 30]);
});
test("returns nothing without the marker", () => {
expect(candidateNumbers("- #1 - looks like #1", 35)).toEqual([]);
});
});
describe("pendingNotice", () => {
test("waits out the grace period from the latest notice", () => {
const fresh = pendingNotice(issue(35, "t"), [notice([10], daysAgo(2.9))], config);
expect(fresh.kind).toBe("skip");
const aged = pendingNotice(issue(35, "t"), [notice([10], daysAgo(3.1))], config);
expect(aged.kind).toBe("pending");
const reposted = pendingNotice(
issue(35, "t"),
[notice([10], daysAgo(6)), notice([10], daysAgo(1), { id: 902 })],
config,
);
expect(reposted.kind).toBe("skip");
});
test("an objection posted before a re-posted notice still keeps the issue open", () => {
const verdict = pendingNotice(
issue(35, "t"),
[notice([10], daysAgo(10)), humanComment(daysAgo(7)), notice([10], daysAgo(4), { id: 902 })],
config,
);
expect(verdict).toEqual({ kind: "skip", reason: "someone replied after the notice" });
});
test("a zero-day grace period acts on the notice at once", () => {
const verdict = pendingNotice(issue(35, "t"), [notice([10], daysAgo(0.01))], { ...config, graceDays: 0 });
expect(verdict.kind).toBe("pending");
});
test("a human reply after the notice keeps the issue open, a bot reply does not", () => {
const human = pendingNotice(issue(35, "t"), [notice([10], daysAgo(5)), humanComment(daysAgo(4))], config);
expect(human).toEqual({ kind: "skip", reason: "someone replied after the notice" });
const bot = pendingNotice(
issue(35, "t"),
[notice([10], daysAgo(5)), { id: 903, body: "triage", created_at: daysAgo(4), user: { type: "Bot", login: "triage[bot]" } }],
config,
);
expect(bot.kind).toBe("pending");
});
test("a human quoting the marker is not a notice", () => {
const quoted = pendingNotice(issue(35, "t"), [notice([10], daysAgo(5), { user: { type: "User", login: "reporter" } })], config);
expect(quoted).toEqual({ kind: "skip", reason: "carries no duplicate notice" });
});
test("never closes an issue twice: a reopened issue is left alone", () => {
const reopened = pendingNotice(
issue(35, "t"),
[notice([10], daysAgo(9)), { id: 904, body: `Closed automatically\n\n${CLOSED_MARKER}`, created_at: daysAgo(5), user: { type: "Bot", login: "github-actions[bot]" } }],
config,
);
expect(reopened).toEqual({ kind: "skip", reason: "was reopened after an automatic close" });
});
test("skips pull requests and issues whose only candidates are newer", () => {
expect(pendingNotice(issue(35, "t", { pull_request: {} }), [notice([10], daysAgo(5))], config).kind).toBe("skip");
expect(pendingNotice(issue(35, "t"), [notice([40], daysAgo(5))], config)).toEqual({
kind: "skip",
reason: "no candidate is older than this issue",
});
});
});
describe("duplicateTarget", () => {
const reporter = issue(35, "[Bug]: Gemma 4-e4b fails on Vertex");
test("closes only against the earliest open issue with the identical normalized title", () => {
const verdict = duplicateTarget(
reporter,
[issue(10, "[Bug]: Gemma 4-e4n fails on Vertex"), issue(20, "[bug]: gemma 4-e4b fails on vertex"), issue(30, "[Bug]: Gemma 4-e4b fails on Vertex")],
[],
);
expect(verdict).toEqual({ kind: "close", duplicateOf: 20 });
});
test("a near miss in the title is not a duplicate", () => {
const verdict = duplicateTarget(reporter, [issue(10, "[Bug]: Gemma 4-e4n fails on Vertex")], []);
expect(verdict).toEqual({ kind: "skip", reason: "no older open issue has the identical title" });
});
test("bare template titles never match each other", () => {
const verdict = duplicateTarget(issue(35, "[Bug]: "), [issue(10, "[Bug]: ")], []);
expect(verdict.kind).toBe("skip");
expect(verdict.kind === "skip" && verdict.reason).toContain("too short");
});
test("a closed candidate or a pull request is never the target", () => {
expect(duplicateTarget(reporter, [issue(10, reporter.title, { state: "closed" })], []).kind).toBe("skip");
expect(duplicateTarget(reporter, [issue(10, reporter.title, { pull_request: {} })], []).kind).toBe("skip");
});
test("a thumbs down on the notice keeps the issue open", () => {
const verdict = duplicateTarget(reporter, [issue(10, reporter.title)], [{ content: "+1" }, { content: "-1" }]);
expect(verdict).toEqual({ kind: "skip", reason: "someone gave the notice a thumbs down" });
});
});
describe("sweepIssue", () => {
const reporter = issue(35, "[Bug]: Gemma 4-e4b fails on Vertex");
const original = issue(10, "[Bug]: Gemma 4-e4b fails on Vertex");
function fakeApi(
comments: readonly Comment[] = [notice([10], daysAgo(5))],
reactionsByNotice: Readonly<Record<number, readonly Reaction[]>> = {},
): { readonly api: GitHubApi; readonly writes: readonly string[] } {
const writes: string[] = [];
const api: GitHubApi = {
request: async <T>(method: string, path: string, body?: object): Promise<T> => {
if (method !== "GET") {
writes.push(`${method} ${path} ${JSON.stringify(body)}`);
return {} as T;
}
if (path.startsWith("/repos/BerriAI/litellm/issues/35/comments")) {
return comments as T;
}
const reactionsPath = path.match(/^\/repos\/BerriAI\/litellm\/issues\/comments\/(\d+)\/reactions/);
if (reactionsPath) {
return (reactionsByNotice[Number(reactionsPath[1])] ?? []) as T;
}
if (path === "/repos/BerriAI/litellm/issues/10") {
return original as T;
}
throw new Error(`unexpected GET ${path}`);
},
};
return { api, writes };
}
test("a dry run reports the close and writes nothing", async () => {
const { api, writes } = fakeApi();
const verdict = await sweepIssue(api, { ...config, dryRun: true }, reporter);
expect(verdict).toEqual({ kind: "close", duplicateOf: 10 });
expect(writes).toEqual([]);
});
test("a thumbs down on an earlier notice still keeps the issue open", async () => {
const { api, writes } = fakeApi([notice([10], daysAgo(9)), notice([10], daysAgo(5), { id: 902 })], { 900: [{ content: "-1" }] });
const verdict = await sweepIssue(api, config, reporter);
expect(verdict).toEqual({ kind: "skip", reason: "someone gave the notice a thumbs down" });
expect(writes).toEqual([]);
});
test("a real run comments, labels, then closes with the duplicate reason", async () => {
const { api, writes } = fakeApi();
const verdict = await sweepIssue(api, config, reporter);
expect(verdict).toEqual({ kind: "close", duplicateOf: 10 });
expect(writes.map((write) => write.split(" ").slice(0, 2).join(" "))).toEqual([
"POST /repos/BerriAI/litellm/issues/35/comments",
"POST /repos/BerriAI/litellm/issues/35/labels",
"PATCH /repos/BerriAI/litellm/issues/35",
]);
expect(writes[0]).toContain("duplicate of #10");
expect(writes[0]).toContain("unanswered for 3 days");
expect(writes[0]).toContain(CLOSED_MARKER);
expect(writes[1]).toContain('{"labels":["duplicate"]}');
expect(writes[2]).toContain('{"state":"closed","state_reason":"duplicate"}');
});
});
describe("reopenTarget", () => {
const closedByBot = (overrides: Partial<Issue> = {}): Issue =>
issue(35, "t", { state: "closed", closed_by: { type: "Bot" }, ...overrides });
const closeMarker = (createdAt: string): Comment => ({
id: 905,
body: `Closed automatically as a duplicate of #10.\n\n${CLOSED_MARKER}`,
created_at: createdAt,
user: { type: "Bot", login: "github-actions[bot]" },
});
test("a reporter reply after the automatic close reopens", () => {
const verdict = reopenTarget(closedByBot(), [closeMarker(daysAgo(2)), humanComment(daysAgo(1))]);
expect(verdict).toEqual({ kind: "reopen" });
});
test("an issue closed by a person stays closed", () => {
const verdict = reopenTarget(closedByBot({ closed_by: { type: "User" } }), [
closeMarker(daysAgo(2)),
humanComment(daysAgo(1)),
]);
expect(verdict).toEqual({ kind: "skip", reason: "was closed by a person" });
});
test("without the automatic-close marker nothing reopens", () => {
const verdict = reopenTarget(closedByBot(), [humanComment(daysAgo(1))]);
expect(verdict).toEqual({ kind: "skip", reason: "carries no automatic-close marker" });
});
test("a maintainer reply alone does not reopen", () => {
const verdict = reopenTarget(closedByBot(), [
closeMarker(daysAgo(2)),
humanComment(daysAgo(1), "Confirmed duplicate", "maintainer"),
]);
expect(verdict).toEqual({ kind: "skip", reason: "the reporter has not replied since the close" });
});
test("a reporter comment from before the close does not reopen", () => {
const verdict = reopenTarget(closedByBot(), [humanComment(daysAgo(3)), closeMarker(daysAgo(2))]);
expect(verdict).toEqual({ kind: "skip", reason: "the reporter has not replied since the close" });
});
test("a pull request never reopens", () => {
const verdict = reopenTarget(closedByBot({ pull_request: {} }), [closeMarker(daysAgo(2)), humanComment(daysAgo(1))]);
expect(verdict).toEqual({ kind: "skip", reason: "is a pull request" });
});
});
describe("sweepClosedIssue", () => {
function fakeApi(issueBody: Issue, comments: readonly Comment[]): { readonly api: GitHubApi; readonly writes: readonly string[] } {
const writes: string[] = [];
const api: GitHubApi = {
request: async <T>(method: string, path: string, body?: object): Promise<T> => {
if (method !== "GET") {
writes.push(`${method} ${path} ${JSON.stringify(body)}`);
return {} as T;
}
if (path.startsWith("/repos/BerriAI/litellm/issues/35/comments")) {
return comments as T;
}
if (path === "/repos/BerriAI/litellm/issues/35") {
return issueBody as T;
}
throw new Error(`unexpected GET ${path}`);
},
};
return { api, writes };
}
const closedByBot = issue(35, "t", { state: "closed", closed_by: { type: "Bot" } });
const closeMarker: Comment = {
id: 905,
body: `Closed automatically as a duplicate of #10.\n\n${CLOSED_MARKER}`,
created_at: daysAgo(2),
user: { type: "Bot", login: "github-actions[bot]" },
};
test("a real run unlabels, reopens, then explains", async () => {
const { api, writes } = fakeApi(closedByBot, [closeMarker, humanComment(daysAgo(1))]);
const verdict = await sweepClosedIssue(api, config, 35);
expect(verdict).toEqual({ kind: "reopen" });
expect(writes).toEqual([
"DELETE /repos/BerriAI/litellm/issues/35/labels/duplicate undefined",
'PATCH /repos/BerriAI/litellm/issues/35 {"state":"open"}',
`POST /repos/BerriAI/litellm/issues/35/comments {"body":"${REOPEN_COMMENT}"}`,
]);
});
test("a dry run reports the reopen and writes nothing", async () => {
const { api, writes } = fakeApi(closedByBot, [closeMarker, humanComment(daysAgo(1))]);
const verdict = await sweepClosedIssue(api, { ...config, dryRun: true }, 35);
expect(verdict).toEqual({ kind: "reopen" });
expect(writes).toEqual([]);
});
});
describe("readConfig", () => {
test("defaults to a real run with a 3-day grace period", () => {
const parsed = readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "BerriAI/litellm" }, NOW);
expect(parsed).toEqual({ token: "t", repo: "BerriAI/litellm", graceDays: 3, dryRun: false, now: NOW });
});
test("honors DRY_RUN and GRACE_PERIOD_DAYS overrides", () => {
const parsed = readConfig(
{ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "o/r", DRY_RUN: "true", GRACE_PERIOD_DAYS: "0" },
NOW,
);
expect(parsed.dryRun).toBe(true);
expect(parsed.graceDays).toBe(0);
});
test("an empty GRACE_PERIOD_DAYS, as a schedule run renders it, means the default", () => {
const parsed = readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "o/r", GRACE_PERIOD_DAYS: "" }, NOW);
expect(parsed.graceDays).toBe(3);
});
test("refuses a missing token, a malformed repository, or a bad grace period", () => {
expect(() => readConfig({ GITHUB_REPOSITORY: "o/r" }, NOW)).toThrow("GITHUB_TOKEN");
expect(() => readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "litellm" }, NOW)).toThrow("owner/repo");
expect(() => readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "o/r", GRACE_PERIOD_DAYS: "-1" }, NOW)).toThrow(
"GRACE_PERIOD_DAYS",
);
});
});

300
scripts/auto-close-duplicates.ts Normal file
View file

@ -0,0 +1,300 @@
#!/usr/bin/env bun
declare const process: { readonly env: Readonly<Record<string, string | undefined>> };
export interface Issue {
readonly number: number;
readonly title: string;
readonly state: string;
readonly user: { readonly login: string };
readonly closed_by?: { readonly type: string } | null;
readonly pull_request?: unknown;
}
export interface Comment {
readonly id: number;
readonly body: string;
readonly created_at: string;
readonly user: { readonly type: string; readonly login: string };
}
export interface Reaction {
readonly content: string;
}
export interface GitHubApi {
readonly request: <T>(method: "GET" | "POST" | "PATCH" | "DELETE", path: string, body?: object) => Promise<T>;
}
export interface SweepConfig {
readonly repo: string;
readonly graceDays: number;
readonly dryRun: boolean;
readonly now: Date;
}
export type NoticeVerdict =
| { readonly kind: "pending"; readonly notices: readonly Comment[]; readonly candidates: readonly number[] }
| { readonly kind: "skip"; readonly reason: string };
export type CloseVerdict =
| { readonly kind: "close"; readonly duplicateOf: number }
| { readonly kind: "skip"; readonly reason: string };
export type ReopenVerdict =
| { readonly kind: "reopen" }
| { readonly kind: "skip"; readonly reason: string };
export const FLAG_LABEL = "potential-duplicate";
export const CLOSED_MARKER = "<!-- litellm:closed-as-duplicate -->";
export const DEFAULT_GRACE_DAYS = 3;
export const REOPEN_COMMENT =
"Reopened automatically: the reporter replied after the duplicate close, so this needs a human look.";
const NOTICE_MARKER = /<!-- litellm:potential-duplicate candidates=([\d,]*) -->/;
const MIN_TITLE_WORDS = 3;
const PAGE_SIZE = 100;
const DAY_MS = 24 * 60 * 60 * 1000;
const REOPEN_LOOKBACK_DAYS = 30;
const skip = (reason: string): { readonly kind: "skip"; readonly reason: string } => ({ kind: "skip", reason });
export function normalizeTitle(title: string): string {
return title
.toLowerCase()
.replace(/^\s*\[[^\]]*\]\s*:?/, "")
.replace(/[^a-z0-9]+/g, " ")
.trim();
}
export function candidateNumbers(noticeBody: string, issueNumber: number): readonly number[] {
const field = noticeBody.match(NOTICE_MARKER);
if (!field) {
return [];
}
const older = field[1]
.split(",")
.filter((value) => value !== "")
.map(Number)
.filter((candidate) => candidate < issueNumber);
return [...new Set(older)].sort((a, b) => a - b);
}
export function pendingNotice(
issue: Issue,
comments: readonly Comment[],
config: Pick<SweepConfig, "graceDays" | "now">,
): NoticeVerdict {
if (issue.pull_request !== undefined) {
return skip("is a pull request");
}
if (comments.some((comment) => comment.body.includes(CLOSED_MARKER))) {
return skip("was reopened after an automatic close");
}
const notices = comments.filter((comment) => comment.user.type === "Bot" && NOTICE_MARKER.test(comment.body));
const first = notices[0];
const latest = notices[notices.length - 1];
if (first === undefined || latest === undefined) {
return skip("carries no duplicate notice");
}
const ageDays = (config.now.getTime() - new Date(latest.created_at).getTime()) / DAY_MS;
if (ageDays < config.graceDays) {
return skip(`notice is ${ageDays.toFixed(1)} days old, grace period is ${config.graceDays}`);
}
const firstNoticeAt = new Date(first.created_at);
if (comments.some((comment) => comment.user.type !== "Bot" && new Date(comment.created_at) > firstNoticeAt)) {
return skip("someone replied after the notice");
}
const candidates = candidateNumbers(latest.body, issue.number);
if (candidates.length === 0) {
return skip("no candidate is older than this issue");
}
return { kind: "pending", notices, candidates };
}
export function duplicateTarget(
issue: Issue,
candidates: readonly Issue[],
reactions: readonly Reaction[],
): CloseVerdict {
if (reactions.some((reaction) => reaction.content === "-1")) {
return skip("someone gave the notice a thumbs down");
}
const title = normalizeTitle(issue.title);
if (title.split(" ").length < MIN_TITLE_WORDS) {
return skip(`title "${issue.title}" is too short to match on`);
}
const original = candidates.find(
(candidate) =>
candidate.state === "open" && candidate.pull_request === undefined && normalizeTitle(candidate.title) === title,
);
if (original === undefined) {
return skip("no older open issue has the identical title");
}
return { kind: "close", duplicateOf: original.number };
}
export function reopenTarget(issue: Issue, comments: readonly Comment[]): ReopenVerdict {
if (issue.pull_request !== undefined) {
return skip("is a pull request");
}
if (issue.closed_by?.type !== "Bot") {
return skip("was closed by a person");
}
const marker = comments.find((comment) => comment.body.includes(CLOSED_MARKER));
if (marker === undefined) {
return skip("carries no automatic-close marker");
}
const markerAt = new Date(marker.created_at);
if (!comments.some((comment) => comment.user.login === issue.user.login && new Date(comment.created_at) > markerAt)) {
return skip("the reporter has not replied since the close");
}
return { kind: "reopen" };
}
export function closingComment(duplicateOf: number, graceDays: number): string {
return `Closed automatically as a duplicate of #${duplicateOf}. Its title is identical to that older open issue and the duplicate notice above went unanswered for ${graceDays} days. If this is wrong, comment here with how it differs from #${duplicateOf} and this issue will be reopened automatically within a day.
${CLOSED_MARKER}`;
}
async function listAll<T>(api: GitHubApi, path: string, page = 1): Promise<readonly T[]> {
const separator = path.includes("?") ? "&" : "?";
const batch = await api.request<readonly T[]>("GET", `${path}${separator}per_page=${PAGE_SIZE}&page=${page}`);
return batch.length < PAGE_SIZE ? batch : [...batch, ...(await listAll<T>(api, path, page + 1))];
}
async function closeAsDuplicate(
api: GitHubApi,
config: SweepConfig,
issueNumber: number,
duplicateOf: number,
): Promise<void> {
const issuePath = `/repos/${config.repo}/issues/${issueNumber}`;
await api.request("POST", `${issuePath}/comments`, { body: closingComment(duplicateOf, config.graceDays) });
await api.request("POST", `${issuePath}/labels`, { labels: ["duplicate"] });
await api.request("PATCH", issuePath, { state: "closed", state_reason: "duplicate" });
}
async function reopenForReporter(api: GitHubApi, config: SweepConfig, issueNumber: number): Promise<void> {
const issuePath = `/repos/${config.repo}/issues/${issueNumber}`;
await api.request("DELETE", `${issuePath}/labels/duplicate`);
await api.request("PATCH", issuePath, { state: "open" });
await api.request("POST", `${issuePath}/comments`, { body: REOPEN_COMMENT });
}
export async function sweepClosedIssue(api: GitHubApi, config: SweepConfig, issueNumber: number): Promise<ReopenVerdict> {
const issue = await api.request<Issue>("GET", `/repos/${config.repo}/issues/${issueNumber}`);
const comments = await listAll<Comment>(api, `/repos/${config.repo}/issues/${issueNumber}/comments`);
const verdict = reopenTarget(issue, comments);
if (verdict.kind === "reopen" && !config.dryRun) {
await reopenForReporter(api, config, issueNumber);
}
return verdict;
}
export async function sweepIssue(api: GitHubApi, config: SweepConfig, issue: Issue): Promise<CloseVerdict> {
const comments = await listAll<Comment>(api, `/repos/${config.repo}/issues/${issue.number}/comments`);
const pending = pendingNotice(issue, comments, config);
if (pending.kind === "skip") {
return pending;
}
const reactions = (
await Promise.all(
pending.notices.map((notice) => listAll<Reaction>(api, `/repos/${config.repo}/issues/comments/${notice.id}/reactions`)),
)
).flat();
const candidates = await Promise.all(
pending.candidates.map((candidate) => api.request<Issue>("GET", `/repos/${config.repo}/issues/${candidate}`)),
);
const verdict = duplicateTarget(issue, candidates, reactions);
if (verdict.kind === "close" && !config.dryRun) {
await closeAsDuplicate(api, config, issue.number, verdict.duplicateOf);
}
return verdict;
}
function describe(issue: Issue, verdict: CloseVerdict, dryRun: boolean): string {
if (verdict.kind === "skip") {
return `#${issue.number}: skipped, ${verdict.reason}`;
}
return `#${issue.number}: ${dryRun ? "would close" : "closed"} as a duplicate of #${verdict.duplicateOf}`;
}
export async function sweep(api: GitHubApi, config: SweepConfig): Promise<readonly CloseVerdict[]> {
const issues = await listAll<Issue>(api, `/repos/${config.repo}/issues?state=open&labels=${FLAG_LABEL}`);
console.log(`${issues.length} open issues carry the ${FLAG_LABEL} label in ${config.repo}${config.dryRun ? " (dry run)" : ""}`);
return issues.reduce<Promise<readonly CloseVerdict[]>>(async (previous, issue) => {
const verdicts = await previous;
const verdict = await sweepIssue(api, config, issue);
console.log(describe(issue, verdict, config.dryRun));
return [...verdicts, verdict];
}, Promise.resolve([]));
}
function describeReopen(issueNumber: number, verdict: ReopenVerdict, dryRun: boolean): string {
if (verdict.kind === "skip") {
return `#${issueNumber}: skipped, ${verdict.reason}`;
}
return `#${issueNumber}: ${dryRun ? "would reopen" : "reopened"} for the reporter's reply`;
}
export async function reopenSweep(api: GitHubApi, config: SweepConfig): Promise<readonly ReopenVerdict[]> {
const since = new Date(config.now.getTime() - REOPEN_LOOKBACK_DAYS * DAY_MS).toISOString();
const closedPath = `/repos/${config.repo}/issues?state=closed&labels=duplicate,${FLAG_LABEL}&since=${encodeURIComponent(since)}`;
const issues = await listAll<Issue>(api, closedPath);
console.log(`${issues.length} recently closed issues carry the duplicate and ${FLAG_LABEL} labels in ${config.repo}${config.dryRun ? " (dry run)" : ""}`);
return issues.reduce<Promise<readonly ReopenVerdict[]>>(async (previous, issue) => {
const verdicts = await previous;
const verdict = await sweepClosedIssue(api, config, issue.number);
console.log(describeReopen(issue.number, verdict, config.dryRun));
return [...verdicts, verdict];
}, Promise.resolve([]));
}
export function readConfig(env: Readonly<Record<string, string | undefined>>, now: Date): SweepConfig & { readonly token: string } {
const token = env.GITHUB_TOKEN;
const repo = env.GITHUB_REPOSITORY;
if (!token || !repo || !/^[\w.-]+\/[\w.-]+$/.test(repo)) {
throw new Error("GITHUB_TOKEN and GITHUB_REPOSITORY (owner/repo) are required");
}
const rawGraceDays = env.GRACE_PERIOD_DAYS?.trim();
const graceDays = rawGraceDays === undefined || rawGraceDays === "" ? DEFAULT_GRACE_DAYS : Number(rawGraceDays);
if (!Number.isFinite(graceDays) || graceDays < 0) {
throw new Error(`GRACE_PERIOD_DAYS must be a non-negative number, got "${env.GRACE_PERIOD_DAYS}"`);
}
return { token, repo, graceDays, dryRun: env.DRY_RUN === "true", now };
}
export function githubApi(token: string): GitHubApi {
return {
request: async <T>(method: "GET" | "POST" | "PATCH" | "DELETE", path: string, body?: object): Promise<T> => {
const response = await fetch(`https://api.github.com${path}`, {
method,
headers: {
Authorization: `Bearer ${token}`,
Accept: "application/vnd.github+json",
"X-GitHub-Api-Version": "2022-11-28",
"User-Agent": "litellm-auto-close-duplicates",
...(body === undefined ? {} : { "Content-Type": "application/json" }),
},
body: body === undefined ? undefined : JSON.stringify(body),
});
if (!response.ok) {
throw new Error(`${method} ${path} failed: ${response.status} ${response.statusText}`);
}
return (await response.json()) as T;
},
};
}
if (import.meta.main) {
const { token, ...config } = readConfig(process.env, new Date());
const api = githubApi(token);
const closeVerdicts = await sweep(api, config);
const reopenVerdicts = await reopenSweep(api, config);
const closed = closeVerdicts.filter((verdict) => verdict.kind === "close").length;
const reopened = reopenVerdicts.filter((verdict) => verdict.kind === "reopen").length;
console.log(
`${config.dryRun ? "Would close" : "Closed"} ${closed} of ${closeVerdicts.length} flagged issues, ${config.dryRun ? "would reopen" : "reopened"} ${reopened}`,
);
}

56
tests/proxy_unit_tests/test_proxy_utils.py
View file

@ -1317,6 +1317,62 @@ def test_proxy_config_state_post_init_callback_call(monkeypatch):
assert config["litellm_settings"]["default_team_settings"][0]["team_id"] == "test"
@pytest.mark.asyncio
async def test_default_team_settings_newrelic_resolves_traces_and_metrics():
"""Static `default_team_settings` is the config-file twin of POST /team/callback.
A team pinned to New Relic through `default_team_settings` must reach the
same two loggers the dynamic path does: the per-team metrics logger (cost
and usage) and the trace logger (LLM/agent spans). This proves the static
path resolves both, not just one, so the config-file customer gets the
same per-team routing as the API customer.
"""
from litellm.litellm_core_utils.litellm_logging import Logging
from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
from litellm.proxy.proxy_server import ProxyConfig
pc = ProxyConfig()
pc.config = {
"litellm_settings": {
"default_team_settings": [
{
"team_id": "team-a",
"success_callback": ["newrelic"],
"newrelic_api_key": "team-a-ingest-key",
"newrelic_region": "eu",
}
]
}
}
callback_metadata = LiteLLMProxyRequestSetup.add_team_based_callbacks_from_config(
team_id="team-a",
proxy_config=pc,
)
assert callback_metadata is not None
assert callback_metadata.success_callback == ["newrelic"]
assert callback_metadata.callback_vars == {
"newrelic_api_key": "team-a-ingest-key",
"newrelic_region": "eu",
}
logging_obj = Logging(
model="gpt-3.5-turbo",
messages=[{"role": "user", "content": "hi"}],
stream=False,
call_type="completion",
start_time=None,
litellm_call_id="static-nr-1",
function_id="static-nr-1",
)
logging_obj._trusted_callback_vars = tuple(callback_metadata.callback_vars.items())
resolved = logging_obj._resolve_dynamic_callback_string("newrelic")
resolved_names = {type(logger).__name__ for logger in resolved}
assert resolved_names == {"NewRelicMetricsLogger", "NewRelicLogger"}
def test_proxy_config_state_get_config_state_error():
"""
Ensures that get_config_state does not raise an error when the config is not a valid dictionary

55
tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py
View file

@ -12,7 +12,7 @@ import litellm
from litellm.caching.caching import DualCache
from litellm.integrations.SlackAlerting.slack_alerting import SlackAlerting
from litellm.proxy._types import CallInfo, Litellm_EntityType
from litellm.types.integrations.slack_alerting import SlackAlertingCacheKeys
from litellm.types.integrations.slack_alerting import AlertType, SlackAlertingCacheKeys
class TestSlackAlerting(unittest.TestCase):
@ -366,3 +366,56 @@ async def test_scheduled_daily_report_threads_the_pod_lock_manager_through():
_, kwargs = slack_alerting._run_scheduler_helper.await_args
assert kwargs["pod_lock_manager"] is pod_lock_manager
def _slack_alerting_with_env_resolution() -> SlackAlerting:
slack_alerting: Final = SlackAlerting(alerting=["slack"], internal_usage_cache=DualCache())
slack_alerting.periodic_started = True
return slack_alerting
@pytest.mark.asyncio
async def test_send_alert_falls_back_to_alerting_webhook_url_env(monkeypatch):
monkeypatch.delenv("SLACK_WEBHOOK_URL", raising=False)
monkeypatch.setenv("ALERTING_WEBHOOK_URL", "https://chat.example.com/hooks/abc")
slack_alerting: Final = _slack_alerting_with_env_resolution()
await slack_alerting.send_alert(
message="budget crossed",
level="High",
alert_type=AlertType.budget_alerts,
alerting_metadata={},
)
assert slack_alerting.log_queue[0]["url"] == "https://chat.example.com/hooks/abc"
@pytest.mark.asyncio
async def test_send_alert_prefers_slack_webhook_url_over_fallback(monkeypatch):
monkeypatch.setenv("SLACK_WEBHOOK_URL", "https://hooks.slack.com/services/T0/B0/X0")
monkeypatch.setenv("ALERTING_WEBHOOK_URL", "https://chat.example.com/hooks/abc")
slack_alerting: Final = _slack_alerting_with_env_resolution()
await slack_alerting.send_alert(
message="budget crossed",
level="High",
alert_type=AlertType.budget_alerts,
alerting_metadata={},
)
assert slack_alerting.log_queue[0]["url"] == "https://hooks.slack.com/services/T0/B0/X0"
@pytest.mark.asyncio
async def test_send_alert_raises_when_no_webhook_url_configured(monkeypatch):
monkeypatch.delenv("SLACK_WEBHOOK_URL", raising=False)
monkeypatch.delenv("ALERTING_WEBHOOK_URL", raising=False)
slack_alerting: Final = _slack_alerting_with_env_resolution()
with pytest.raises(ValueError, match="SLACK_WEBHOOK_URL / ALERTING_WEBHOOK_URL"):
await slack_alerting.send_alert(
message="budget crossed",
level="High",
alert_type=AlertType.budget_alerts,
alerting_metadata={},
)

17
tests/test_litellm/integrations/SlackAlerting/test_slack_alerting_digest.py
View file

@ -79,6 +79,23 @@ class TestDigestMode(unittest.IsolatedAsyncioTestCase):
self.assertEqual(len(self.slack_alerting.digest_buckets), 2)
async def test_digest_falls_back_to_alerting_webhook_url_env(self):
"""With SLACK_WEBHOOK_URL unset, the digest entry resolves ALERTING_WEBHOOK_URL instead."""
env = {k: v for k, v in os.environ.items() if k != "SLACK_WEBHOOK_URL"}
env["ALERTING_WEBHOOK_URL"] = "https://chat.example.com/hooks/abc"
with unittest.mock.patch.dict(os.environ, env, clear=True):
await self.slack_alerting.send_alert(
message="`Requests are hanging`",
level="Medium",
alert_type=AlertType.llm_requests_hanging,
alerting_metadata={},
request_model="gemini-2.5-flash",
api_base="None",
)
bucket = list(self.slack_alerting.digest_buckets.values())[0]
self.assertEqual(bucket["webhook_url"], "https://chat.example.com/hooks/abc")
async def test_non_digest_alert_goes_to_queue(self):
"""Alert types without digest enabled should go straight to the log queue."""
message = "Budget exceeded"

22
tests/test_litellm/integrations/otel/test_otel_v2_components.py
View file

@ -3,6 +3,7 @@ baggage helpers, metrics, the typed coercion helpers, mapper branches, span-name
builders, and the registry validator's failure paths. Needs the OTel SDK."""
import json
from dataclasses import replace
import pytest
@ -215,6 +216,27 @@ def test_genai_mapper_all_request_params():
assert attrs["server.port"] == 443
def test_genai_mapper_cache_token_attrs():
cached = replace(
_full_llm_call(),
usage=LLMUsage(
input_tokens=10,
output_tokens=5,
total_tokens=15,
cache_creation_input_tokens=7,
cache_read_input_tokens=3,
),
)
attrs = GenAIMapper().map(cached)
assert attrs[GenAI.USAGE_CACHE_CREATION_INPUT_TOKENS] == 7
assert attrs[GenAI.USAGE_CACHE_READ_INPUT_TOKENS] == 3
# No cache usage keeps the span sparse: neither key present.
uncached = GenAIMapper().map(_full_llm_call())
assert GenAI.USAGE_CACHE_CREATION_INPUT_TOKENS not in uncached
assert GenAI.USAGE_CACHE_READ_INPUT_TOKENS not in uncached
def test_genai_mapper_stamps_input_output_messages():
data = LLMCallSpanData(
operation=GenAIOperation.CHAT,

22
tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py
View file

@ -525,6 +525,28 @@ def test_llm_call_adapter_extracts_all_fields():
assert data.identity.key_hash == "hsh"
def test_llm_call_adapter_extracts_cache_tokens_from_usage_object():
payload = _sample_payload()
payload["metadata"] = {
**payload["metadata"],
"usage_object": {
"prompt_tokens": 10,
"completion_tokens": 5,
"cache_creation_input_tokens": 7,
"cache_read_input_tokens": 3,
},
}
data = LLMCallSpanData.from_standard_logging_payload(payload)
assert data.usage.cache_creation_input_tokens == 7
assert data.usage.cache_read_input_tokens == 3
def test_llm_call_adapter_cache_tokens_none_without_usage_object():
data = LLMCallSpanData.from_standard_logging_payload(_sample_payload())
assert data.usage.cache_creation_input_tokens is None
assert data.usage.cache_read_input_tokens is None
def test_llm_call_adapter_failure_path():
payload = _sample_payload(
status="failure",

35
tests/test_litellm/litellm_core_utils/test_token_counter.py
View file

@ -1377,3 +1377,38 @@ def test_anthropic_document_title_and_context_add_their_tokens():
{"type": "document", "source": source},
]
)
def test_openai_file_block_prices_like_the_equivalent_anthropic_document():
"""An inline `file` is a `document` in the chat-completions dialect, so it must price identically, not raise.
Before the fix `file` was missing from the content-block match even though `ChatCompletionFileObject`
is in the union this counter accepts, so every local count of a Responses `input_file` raised
`Invalid content item type: file` and surfaced as a 500 on /v1/responses/input_tokens.
"""
prompt = {"type": "text", "text": "Summarize this file."}
inline_file = {
"type": "file",
"file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQK"},
}
document = {
"type": "document",
"title": "report.pdf",
"source": {"type": "base64", "media_type": "application/pdf", "data": "JVBERi0xLjQK"},
}
assert _count_user_content([prompt, inline_file]) == _count_user_content([prompt, document])
assert _count_user_content([prompt, inline_file]) > _count_user_content([prompt])
def test_openai_file_block_without_inline_bytes_counts_what_it_carries():
"""A `file` block naming an uploaded file has no bytes to price, so it adds only the filename's tokens."""
prompt = {"type": "text", "text": "Summarize this file."}
by_id = {"type": "file", "file": {"file_id": "file-abc123"}}
assert _count_user_content([prompt, by_id]) == _count_user_content([prompt])
named = {"type": "file", "file": {"file_id": "file-abc123", "filename": "report.pdf"}}
assert _count_user_content([prompt, named]) == _count_user_content(
[prompt, {"type": "text", "text": "report.pdf"}]
)

30
tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py
View file

@ -957,6 +957,28 @@ def test_transform_request_helper_includes_anthropic_beta_and_tools():
assert fields["tools"][0]["type"] == "computer_20250124"
def test_config_blocks_do_not_leak_into_inference_config():
"""Regression: inferenceConfig was built before the config blocks were popped, so a dead
nested copy of each block (guardrailConfig, performanceConfig, serviceTier) rode inside
inferenceConfig alongside the real top-level one."""
data = AmazonConverseConfig()._transform_request_helper(
model="anthropic.claude-haiku-4-5-20251001-v1:0",
system_content_blocks=[],
optional_params={
"maxTokens": 100,
"guardrailConfig": {"guardrailIdentifier": "gr-id", "guardrailVersion": "DRAFT"},
"performanceConfig": {"latency": "optimized"},
"serviceTier": {"type": "priority"},
},
messages=[{"role": "user", "content": "hi"}],
)
assert data["inferenceConfig"] == {"maxTokens": 100}
assert data["guardrailConfig"] == {"guardrailIdentifier": "gr-id", "guardrailVersion": "DRAFT"}
assert data["performanceConfig"] == {"latency": "optimized"}
assert data["serviceTier"] == {"type": "priority"}
def test_parallel_tool_calls_config_kept_for_sonnet_5(monkeypatch):
old_env = os.environ.get("LITELLM_LOCAL_MODEL_COST_MAP")
old_cost = litellm.model_cost
@ -2853,17 +2875,11 @@ def test_guarded_text_guardrail_config_preserved():
headers={},
)
# GuardrailConfig should be present at top level
assert "guardrailConfig" in result
assert result["guardrailConfig"]["guardrailIdentifier"] == "gr-abc123"
# GuardrailConfig should also be in inferenceConfig
assert "inferenceConfig" in result
assert "guardrailConfig" in result["inferenceConfig"]
assert (
result["inferenceConfig"]["guardrailConfig"]["guardrailIdentifier"]
== "gr-abc123"
)
assert "guardrailConfig" not in result["inferenceConfig"]
def test_auto_convert_last_user_message_to_guarded_text():

234
tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
View file

@ -163,6 +163,240 @@ def test_messages_to_responses_input_with_tool():
}
def test_messages_to_responses_input_preserves_images():
"""An image block must survive the round trip, or OpenAI counts only the text.
A 256x256 image is worth 255 tokens to OpenAI's counting API; dropping it
turned a 268-token request into a 13-token one.
"""
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "What is in this image?"},
{
"type": "image_url",
"image_url": {"url": "data:image/png;base64,iVBORw0KGgo=", "detail": "high"},
},
],
}
]
input_items, instructions = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert instructions is None
assert input_items == [
{
"role": "user",
"content": (
{"type": "input_text", "text": "What is in this image?"},
{
"type": "input_image",
"image_url": "data:image/png;base64,iVBORw0KGgo=",
"detail": "high",
},
),
}
]
def test_messages_to_responses_input_image_without_detail_defaults_to_auto():
messages = [
{
"role": "user",
"content": [{"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}}],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items[0]["content"] == (
{"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
)
def test_messages_to_responses_input_bare_string_image_url_is_preserved():
messages = [{"role": "user", "content": [{"type": "image_url", "image_url": "https://example.com/cat.png"}]}]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items[0]["content"] == (
{"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
)
def test_messages_to_responses_input_text_only_blocks_stay_a_joined_string():
"""Text-only content must keep collapsing to a string so existing counts do not shift."""
messages = [
{
"role": "user",
"content": [{"type": "text", "text": "first"}, {"type": "text", "text": "second"}],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [{"role": "user", "content": "first\nsecond"}]
def test_messages_to_responses_input_drops_unmappable_blocks():
"""A block with no Responses API equivalent is skipped, never forwarded verbatim."""
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "hi"},
{"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
{"type": "input_audio", "input_audio": {"data": "AAAA", "format": "wav"}},
],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items[0]["content"] == (
{"type": "input_text", "text": "hi"},
{"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
)
def test_messages_to_responses_input_assistant_blocks_collapse_to_a_string():
"""An assistant turn must never forward chat `text` blocks.
The Responses API only accepts output_text and refusal inside an assistant turn, so
forwarding them 400s the whole request and silently drops the count back to the local
tokenizer, which is exactly what defeats the image fix above.
"""
messages = [
{"role": "user", "content": [{"type": "text", "text": "What is the capital of France?"}]},
{"role": "assistant", "content": [{"type": "text", "text": "Paris."}]},
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [
{"role": "user", "content": "What is the capital of France?"},
{"role": "assistant", "content": "Paris."},
]
def test_messages_to_responses_input_assistant_image_block_is_dropped():
"""An image part is illegal inside an assistant turn, so it must not reach the provider."""
messages = [
{
"role": "assistant",
"content": [
{"type": "text", "text": "Here it is"},
{"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [{"role": "assistant", "content": "Here it is"}]
def test_messages_to_responses_input_keeps_user_image_alongside_an_assistant_turn():
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "What is in this image?"},
{"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
],
},
{"role": "assistant", "content": [{"type": "text", "text": "A cat."}]},
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [
{
"role": "user",
"content": (
{"type": "input_text", "text": "What is in this image?"},
{"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
),
},
{"role": "assistant", "content": "A cat."},
]
def test_messages_to_responses_input_preserves_inline_files():
"""An inline file must survive the round trip, or the count silently drops the file.
A small PDF is worth 36 tokens to OpenAI's counting API; dropping it left the same
request counting 13, the text-only total.
"""
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "Summarize this file."},
{
"type": "file",
"file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0="},
},
],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [
{
"role": "user",
"content": (
{"type": "input_text", "text": "Summarize this file."},
{
"type": "input_file",
"filename": "report.pdf",
"file_data": "data:application/pdf;base64,JVBERi0=",
},
),
}
]
def test_messages_to_responses_input_drops_a_file_with_no_inline_data():
"""OpenAI rejects `file_data` without a `filename`, and a rejected request loses the whole count."""
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "Summarize this file."},
{"type": "file", "file": {"file_data": "data:application/pdf;base64,JVBERi0="}},
{"type": "file", "file": {"file_id": "file-abc123"}},
],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [{"role": "user", "content": "Summarize this file."}]
def test_messages_to_responses_input_assistant_file_block_is_dropped():
"""A file part is illegal inside an assistant turn, so it must not reach the provider."""
messages = [
{
"role": "assistant",
"content": [
{"type": "text", "text": "Here it is"},
{
"type": "file",
"file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0="},
},
],
}
]
input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
assert input_items == [{"role": "assistant", "content": "Here it is"}]
def test_validate_request_valid():
"""Test that valid requests pass validation."""
config = OpenAICountTokensConfig()

238
tests/test_litellm/llms/openai/test_openai_workload_identity.py Normal file
View file

@ -0,0 +1,238 @@
import json
import sys
from pathlib import Path
from typing import Final
import httpx
import pytest
import respx
from openai import AsyncOpenAI, OpenAI
import litellm
from litellm.llms.litellm_proxy.responses.transformation import LiteLLMProxyResponsesAPIConfig
from litellm.llms.openai.common_utils import BaseOpenAILLM, OpenAIError
from litellm.llms.openai.openai import OpenAIChatCompletion
from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
from litellm.llms.openai.workload_identity import (
OpenAIWorkloadIdentityConfig,
_workload_identity_auth,
get_workload_identity_bearer_token,
resolve_openai_workload_identity_config,
)
from litellm.types.router import GenericLiteLLMParams
TOKEN_EXCHANGE_URL: Final = "https://auth.openai.com/oauth/token"
@pytest.fixture
def wif_env(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> OpenAIWorkloadIdentityConfig:
token_file: Final = tmp_path / "subject_token.jwt"
token_file.write_text("subject-token-from-file")
monkeypatch.delenv("OPENAI_API_KEY", raising=False)
monkeypatch.delenv("OPENAI_BASE_URL", raising=False)
monkeypatch.delenv("OPENAI_API_BASE", raising=False)
monkeypatch.setattr(litellm, "api_base", None)
monkeypatch.setenv("OPENAI_IDENTITY_PROVIDER_ID", "idp_test123")
monkeypatch.setenv("OPENAI_SERVICE_ACCOUNT_ID", "user-test456")
monkeypatch.setenv("OPENAI_IDENTITY_TOKEN_FILE", str(token_file))
_workload_identity_auth.cache_clear()
litellm.in_memory_llm_clients_cache.flush_cache()
return OpenAIWorkloadIdentityConfig(
identity_provider_id="idp_test123",
service_account_id="user-test456",
token_file=str(token_file),
)
def mock_token_exchange(access_token: str = "exchanged-bearer-token") -> respx.Route:
return respx.post(TOKEN_EXCHANGE_URL).mock(
return_value=httpx.Response(200, json={"access_token": access_token, "expires_in": 3600})
)
class TestResolveConfig:
def test_resolves_from_env(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) == wif_env
def test_static_api_key_wins(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
assert resolve_openai_workload_identity_config(api_key="sk-static", api_base=None) is None
def test_env_openai_api_key_wins(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch
) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-from-env")
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) is None
@pytest.mark.parametrize("empty_key", ["", " "])
def test_empty_api_key_arg_does_not_disable_wif(
self, wif_env: OpenAIWorkloadIdentityConfig, empty_key: str
) -> None:
assert resolve_openai_workload_identity_config(api_key=empty_key, api_base=None) == wif_env
@pytest.mark.parametrize("empty_key", ["", " "])
def test_empty_env_openai_api_key_does_not_disable_wif(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch, empty_key: str
) -> None:
monkeypatch.setenv("OPENAI_API_KEY", empty_key)
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) == wif_env
def test_foreign_api_base_disables(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
assert resolve_openai_workload_identity_config(api_key=None, api_base="https://my-vllm.internal/v1") is None
def test_openai_api_base_allows(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
assert resolve_openai_workload_identity_config(api_key=None, api_base="https://api.openai.com/v1") == wif_env
def test_plaintext_http_api_base_disables(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
assert resolve_openai_workload_identity_config(api_key=None, api_base="http://api.openai.com/v1") is None
def test_foreign_env_base_url_disables(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch
) -> None:
monkeypatch.setenv("OPENAI_BASE_URL", "https://my-vllm.internal/v1")
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) is None
def test_openai_env_base_url_allows(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch
) -> None:
monkeypatch.setenv("OPENAI_BASE_URL", "https://api.openai.com/v1")
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) == wif_env
def test_foreign_litellm_api_base_disables(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch
) -> None:
monkeypatch.setattr(litellm, "api_base", "https://my-vllm.internal/v1")
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) is None
@pytest.mark.parametrize(
"missing_var",
["OPENAI_IDENTITY_PROVIDER_ID", "OPENAI_SERVICE_ACCOUNT_ID", "OPENAI_IDENTITY_TOKEN_FILE"],
)
def test_partial_env_disables(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch, missing_var: str
) -> None:
monkeypatch.delenv(missing_var)
assert resolve_openai_workload_identity_config(api_key=None, api_base=None) is None
class TestTokenExchange:
@respx.mock
def test_exchanges_subject_token_for_bearer(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
route: Final = mock_token_exchange()
assert get_workload_identity_bearer_token(wif_env) == "exchanged-bearer-token"
request_body: Final = json.loads(route.calls.last.request.content)
assert request_body["grant_type"] == "urn:ietf:params:oauth:grant-type:token-exchange"
assert request_body["subject_token"] == "subject-token-from-file"
assert request_body["subject_token_type"] == "urn:ietf:params:oauth:token-type:jwt"
assert request_body["identity_provider_id"] == "idp_test123"
assert request_body["service_account_id"] == "user-test456"
@respx.mock
def test_token_cached_across_mints(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
route: Final = mock_token_exchange()
first: Final = get_workload_identity_bearer_token(wif_env)
second: Final = get_workload_identity_bearer_token(wif_env)
assert first == second == "exchanged-bearer-token"
assert route.call_count == 1
def test_old_sdk_raises_upgrade_error(
self, wif_env: OpenAIWorkloadIdentityConfig, monkeypatch: pytest.MonkeyPatch
) -> None:
import openai as openai_module
monkeypatch.delattr(openai_module, "auth", raising=False)
monkeypatch.setitem(sys.modules, "openai.auth", None)
with pytest.raises(OpenAIError, match=r"openai>=2\.32\.0"):
wif_env.to_sdk_workload_identity()
class TestClientConstruction:
def test_sync_client_uses_workload_identity(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
client: Final = OpenAIChatCompletion()._get_openai_client(is_async=False, api_key=None, api_base=None)
assert isinstance(client, OpenAI)
assert client.api_key == "workload-identity-auth"
assert client._workload_identity_auth is not None
def test_async_client_uses_workload_identity(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
client: Final = OpenAIChatCompletion()._get_openai_client(is_async=True, api_key=None, api_base=None)
assert isinstance(client, AsyncOpenAI)
assert client.api_key == "workload-identity-auth"
assert client._workload_identity_auth is not None
def test_static_key_client_unaffected(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
client: Final = OpenAIChatCompletion()._get_openai_client(is_async=False, api_key="sk-static", api_base=None)
assert isinstance(client, OpenAI)
assert client.api_key == "sk-static"
assert client._workload_identity_auth is None
def test_cache_key_separates_wif_identities(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
other_config: Final = OpenAIWorkloadIdentityConfig(
identity_provider_id="idp_other",
service_account_id="user-other",
token_file=wif_env.token_file,
)
keys: Final = tuple(
BaseOpenAILLM.get_openai_client_cache_key(
client_initialization_params={"api_key": None, "is_async": False, "workload_identity_config": config},
client_type="openai",
)
for config in (wif_env, other_config, None)
)
assert len(set(keys)) == 3
@respx.mock
def test_request_carries_exchanged_bearer(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
mock_token_exchange()
completion_route: Final = respx.post("https://api.openai.com/v1/chat/completions").mock(
return_value=httpx.Response(
200,
json={
"id": "chatcmpl-wif",
"object": "chat.completion",
"created": 1,
"model": "gpt-4o-mini",
"choices": [
{
"index": 0,
"message": {"role": "assistant", "content": "ok"},
"finish_reason": "stop",
}
],
"usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
},
)
)
client = OpenAIChatCompletion()._get_openai_client(is_async=False, api_key=None, api_base=None)
assert isinstance(client, OpenAI)
client.chat.completions.create(model="gpt-4o-mini", messages=[{"role": "user", "content": "hi"}])
auth_header: Final = completion_route.calls.last.request.headers["Authorization"]
assert auth_header == "Bearer exchanged-bearer-token"
class TestResponsesValidateEnvironment:
@respx.mock
def test_mints_bearer_when_wif_configured(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
mock_token_exchange()
headers: Final = OpenAIResponsesAPIConfig().validate_environment(
headers={}, model="gpt-4o-mini", litellm_params=GenericLiteLLMParams()
)
assert headers["Authorization"] == "Bearer exchanged-bearer-token"
def test_static_key_wins(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
headers: Final = OpenAIResponsesAPIConfig().validate_environment(
headers={}, model="gpt-4o-mini", litellm_params=GenericLiteLLMParams(api_key="sk-responses")
)
assert headers["Authorization"] == "Bearer sk-responses"
def test_foreign_api_base_skips_wif(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
headers: Final = OpenAIResponsesAPIConfig().validate_environment(
headers={},
model="gpt-4o-mini",
litellm_params=GenericLiteLLMParams(api_base="https://my-vllm.internal/v1"),
)
assert headers["Authorization"] == "Bearer None"
def test_litellm_proxy_subclass_never_mints_wif(self, wif_env: OpenAIWorkloadIdentityConfig) -> None:
headers: Final = LiteLLMProxyResponsesAPIConfig().validate_environment(
headers={}, model="gpt-4o-mini", litellm_params=GenericLiteLLMParams()
)
assert headers["Authorization"] == "Bearer None"

110
tests/test_litellm/llms/vertex_ai/test_vertex_llm_base.py
View file

@ -982,6 +982,116 @@ class TestVertexBase:
assert result_url == f"{gateway_api_base}:embedContent"
def test_check_custom_proxy_vertex_api_base_with_version_path_grafts_default_path(self):
vertex_base = VertexBase()
_, result_url = vertex_base._check_custom_proxy(
api_base="https://aiplatform.googleapis.com/v1beta1",
custom_llm_provider="vertex_ai",
gemini_api_key=None,
endpoint="generateContent",
stream=None,
auth_header="Bearer token123",
url="https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent",
model="gemini-3.5-flash-lite",
)
assert (
result_url
== "https://aiplatform.googleapis.com/v1beta1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent"
)
def test_check_custom_proxy_vertex_api_base_with_version_path_trailing_slash_grafts_default_path(self):
vertex_base = VertexBase()
_, result_url = vertex_base._check_custom_proxy(
api_base="https://internal-gateway.example.com/v1/",
custom_llm_provider="vertex_ai",
gemini_api_key=None,
endpoint="generateContent",
stream=None,
auth_header="Bearer token123",
url="https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent",
model="gemini-3.5-flash-lite",
)
assert (
result_url
== "https://internal-gateway.example.com/v1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent"
)
def test_check_custom_proxy_vertex_api_base_with_version_path_and_query_grafts_before_query(self):
vertex_base = VertexBase()
_, result_url = vertex_base._check_custom_proxy(
api_base="https://internal-gateway.example.com/v1beta1?key=abc",
custom_llm_provider="vertex_ai",
gemini_api_key=None,
endpoint="generateContent",
stream=None,
auth_header="Bearer token123",
url="https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent",
model="gemini-3.5-flash-lite",
)
assert (
result_url
== "https://internal-gateway.example.com/v1beta1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent?key=abc"
)
def test_check_custom_proxy_vertex_api_base_with_version_path_and_query_streaming_appends_alt_sse(self):
vertex_base = VertexBase()
_, result_url = vertex_base._check_custom_proxy(
api_base="https://internal-gateway.example.com/v1beta1?key=abc",
custom_llm_provider="vertex_ai",
gemini_api_key=None,
endpoint="streamGenerateContent",
stream=True,
auth_header="Bearer token123",
url="https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:streamGenerateContent",
model="gemini-3.5-flash-lite",
)
assert (
result_url
== "https://internal-gateway.example.com/v1beta1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:streamGenerateContent?key=abc&alt=sse"
)
def test_check_custom_proxy_vertex_api_base_with_non_version_path_keeps_endpoint_append(self):
vertex_base = VertexBase()
gateway_api_base = "https://gateway.example.com/vertex-proxy"
_, result_url = vertex_base._check_custom_proxy(
api_base=gateway_api_base,
custom_llm_provider="vertex_ai",
gemini_api_key=None,
endpoint="generateContent",
stream=None,
auth_header="Bearer token123",
url="https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/publishers/google/models/gemini-3.5-flash-lite:generateContent",
model="gemini-3.5-flash-lite",
)
assert result_url == f"{gateway_api_base}:generateContent"
def test_check_custom_proxy_vertex_api_base_without_projects_in_default_url_keeps_endpoint_append(self):
vertex_base = VertexBase()
gemma_api_base = "https://example.com/custom/gemma-deployment"
_, result_url = vertex_base._check_custom_proxy(
api_base=gemma_api_base,
custom_llm_provider="vertex_ai",
gemini_api_key=None,
endpoint="predict",
stream=False,
auth_header=None,
url=gemma_api_base,
model="gemma-3-27b-it",
)
assert result_url == f"{gemma_api_base}:predict"
def test_check_custom_proxy_vertex_bare_host_streaming_keeps_single_alt_sse(self):
vertex_base = VertexBase()

1
tests/test_litellm/proxy/client/cli/autoroute/test_commands.py
View file

@ -157,6 +157,7 @@ class TestUpCommand:
assert captured["settings"]["theme"] == "dark"
assert captured["settings"]["env"]["ANTHROPIC_BASE_URL"] == "http://127.0.0.1:5483"
assert captured["settings"]["env"]["ANTHROPIC_AUTH_TOKEN"] == "fixed-master-key"
assert captured["settings"]["env"]["ENABLE_TOOL_SEARCH"] == "true"
assert "apiKeyHelper" not in captured["settings"]
assert captured["settings_mode"] == 0o600

7
tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
View file

@ -19,6 +19,13 @@ def test_sets_base_url_and_auth_token():
merged = merge_claude_settings_static_token({}, "http://127.0.0.1:4000/", "token-abc")
assert merged["env"]["ANTHROPIC_BASE_URL"] == "http://127.0.0.1:4000"
assert merged["env"]["ANTHROPIC_AUTH_TOKEN"] == "token-abc"
assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
def test_preserves_existing_tool_search():
settings = {"env": {"ENABLE_TOOL_SEARCH": "false"}}
merged = merge_claude_settings_static_token(settings, "http://127.0.0.1:4000", "token-abc")
assert merged["env"]["ENABLE_TOOL_SEARCH"] == "false"
def test_drops_stray_api_key():

14
tests/test_litellm/proxy/client/cli/test_agents.py
View file

@ -77,9 +77,19 @@ class TestBuildAgentEnv:
)
assert env["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
assert env["ANTHROPIC_AUTH_TOKEN"] == "sk-key"
assert env["ENABLE_TOOL_SEARCH"] == "true"
assert "OPENAI_BASE_URL" not in env
assert "OPENAI_API_KEY" not in env
def test_anthropic_profile_preserves_existing_tool_search(self):
env = build_agent_env(
{"ENABLE_TOOL_SEARCH": "false"},
"http://localhost:4000",
"sk-key",
frozenset({"anthropic"}),
)
assert env["ENABLE_TOOL_SEARCH"] == "false"
def test_anthropic_profile_drops_existing_api_key(self):
env = build_agent_env(
{"ANTHROPIC_API_KEY": "real-key"},
@ -96,6 +106,7 @@ class TestBuildAgentEnv:
assert env["OPENAI_BASE_URL"] == "http://localhost:4000/v1"
assert env["OPENAI_API_KEY"] == "sk-key"
assert "ANTHROPIC_BASE_URL" not in env
assert "ENABLE_TOOL_SEARCH" not in env
def test_both_profiles_set_everything(self):
env = build_agent_env(
@ -105,6 +116,7 @@ class TestBuildAgentEnv:
assert env["OPENAI_BASE_URL"] == "http://localhost:4000/v1"
assert env["ANTHROPIC_AUTH_TOKEN"] == "sk-key"
assert env["OPENAI_API_KEY"] == "sk-key"
assert env["ENABLE_TOOL_SEARCH"] == "true"
def test_preserves_unrelated_env_and_does_not_mutate_input(self):
base = {"PATH": "/usr/bin", "ANTHROPIC_API_KEY": "real-key"}
@ -201,6 +213,7 @@ class TestRunAgent:
env = calls["env"]
assert env["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
assert env["ANTHROPIC_AUTH_TOKEN"] == "sk-key"
assert env["ENABLE_TOOL_SEARCH"] == "true"
assert "ANTHROPIC_API_KEY" not in env
assert "OPENAI_BASE_URL" not in env
@ -218,6 +231,7 @@ class TestRunAgent:
assert calls["env"]["OPENAI_BASE_URL"] == "http://localhost:4000/v1"
assert calls["env"]["OPENAI_API_KEY"] == "sk-key"
assert "ANTHROPIC_BASE_URL" not in calls["env"]
assert "ENABLE_TOOL_SEARCH" not in calls["env"]
def test_codex_injects_proxy_provider_args_before_user_args(self):
calls = {}

1
tests/test_litellm/proxy/client/cli/test_auth_commands.py
View file

@ -1373,6 +1373,7 @@ class TestLoginConfigClaude:
assert result.exit_code == 0
written = json.loads(settings_path.read_text())
assert written["env"]["ANTHROPIC_BASE_URL"] == "https://test.example.com"
assert written["env"]["ENABLE_TOOL_SEARCH"] == "true"
assert written["apiKeyHelper"] == "/usr/local/bin/lite --base-url https://test.example.com auth print-token"
assert "Configured Claude Code" in result.output

1
tests/test_litellm/proxy/client/cli/test_claude_settings.py
View file

@ -48,6 +48,7 @@ class TestWriteClaudeSettings:
written = json.loads(settings_path.read_text())
assert written["env"]["ANTHROPIC_BASE_URL"] == "https://proxy.example.com"
assert written["env"]["ENABLE_TOOL_SEARCH"] == "true"
assert written["apiKeyHelper"] == "/usr/local/bin/lite --base-url https://proxy.example.com auth print-token"
def test_updates_an_existing_file_preserving_unrelated_settings(self, paths, lite_on_path):

12
tests/test_litellm/proxy/client/cli/test_up_commands.py
View file

@ -55,8 +55,14 @@ class TestMergeClaudeSettings:
}
merged = merge_claude_settings(settings, "http://localhost:4000/", "new-helper")
assert merged["env"]["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
assert merged["apiKeyHelper"] == "new-helper"
def test_preserves_existing_tool_search(self):
settings = {"env": {"ENABLE_TOOL_SEARCH": "false"}}
merged = merge_claude_settings(settings, "http://localhost:4000", "helper")
assert merged["env"]["ENABLE_TOOL_SEARCH"] == "false"
def test_drops_stray_api_key(self):
settings = {"env": {"ANTHROPIC_API_KEY": "leaked-key"}}
merged = merge_claude_settings(settings, "http://localhost:4000", "helper")
@ -64,7 +70,10 @@ class TestMergeClaudeSettings:
def test_works_from_empty_settings(self):
merged = merge_claude_settings({}, "http://localhost:4000", "helper")
assert merged["env"] == {"ANTHROPIC_BASE_URL": "http://localhost:4000"}
assert merged["env"] == {
"ANTHROPIC_BASE_URL": "http://localhost:4000",
"ENABLE_TOOL_SEARCH": "true",
}
assert merged["apiKeyHelper"] == "helper"
def test_does_not_mutate_input(self):
@ -486,6 +495,7 @@ class TestUpCommand:
assert captured["backup_existed"] is True
assert captured["settings"]["theme"] == "dark"
assert captured["settings"]["env"]["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
assert captured["settings"]["env"]["ENABLE_TOOL_SEARCH"] == "true"
assert captured["settings"]["apiKeyHelper"] == "/usr/local/bin/lite auth print-token"
assert json.loads(settings_path.read_text()) == original
assert not backup_path.exists()

32
tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_moderations.py
View file

@ -482,23 +482,25 @@ async def test_openai_moderation_guardrail_streaming_harmful_content():
"metadata": {"guardrails": ["test-openai-moderation"]},
}
# Should raise HTTPException when processing streaming harmful content
from fastapi import HTTPException
# Chunks have already been flushed by end-of-stream moderation, so
# the block surfaces as the in-stream error frame, not a raise.
import json as _json
async def _drain():
result_chunks = []
async for chunk in unified_guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=user_api_key_dict,
response=mock_stream(),
request_data=request_data,
):
result_chunks.append(chunk)
result_chunks = []
async for chunk in unified_guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=user_api_key_dict,
response=mock_stream(),
request_data=request_data,
):
result_chunks.append(chunk)
with pytest.raises(HTTPException) as exc_info:
await _drain()
assert exc_info.value.status_code == 400
assert "Violated OpenAI moderation policy" in str(exc_info.value.detail)
frame = result_chunks[-1]
assert isinstance(frame, bytes)
text = frame.decode()
assert text.startswith("data: ")
assert "Violated OpenAI moderation policy" in text
payload = _json.loads(text[len("data: ") :])
assert payload["error"]["code"] == "400"
@pytest.mark.asyncio

32
tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_openai_moderation_streaming.py
View file

@ -161,19 +161,27 @@ async def test_openai_moderation_guardrail_streaming_harmful_content():
"metadata": {"guardrails": ["test-openai-moderation"]},
}
# Should raise HTTPException
with pytest.raises(HTTPException) as exc_info:
async for (
_
) in unified_guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=user_api_key_dict,
response=mock_stream(),
request_data=request_data,
):
pass
# Chunks have already been flushed by end-of-stream moderation, so
# the block surfaces as the in-stream error frame, not a raise.
import json as _json
assert exc_info.value.status_code == 400
assert "Violated OpenAI moderation policy" in str(exc_info.value.detail)
collected = []
async for (
chunk
) in unified_guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=user_api_key_dict,
response=mock_stream(),
request_data=request_data,
):
collected.append(chunk)
frame = collected[-1]
assert isinstance(frame, bytes)
text = frame.decode()
assert text.startswith("data: ")
assert "Violated OpenAI moderation policy" in text
payload = _json.loads(text[len("data: ") :])
assert payload["error"]["code"] == "400"
@pytest.mark.asyncio

245
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py
View file

@ -5345,3 +5345,248 @@ def test_initialize_bedrock_forwards_aws_external_id():
assert guardrail.optional_params["aws_external_id"] == "external-id-123"
finally:
litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, guardrail)
def _chat_chunk(content: str, finish_reason: str | None) -> litellm.ModelResponseStream:
return litellm.ModelResponseStream(
id="tid",
choices=[
litellm.types.utils.StreamingChoices(
delta=litellm.types.utils.Delta(content=content, role="assistant"),
finish_reason=finish_reason,
index=0,
)
],
created=1,
model="gpt-4o-mini",
object="chat.completion.chunk",
)
def _streaming_litellm_params(**extras):
from litellm.types.guardrails import LitellmParams
return LitellmParams(
guardrail="bedrock",
mode="post_call",
guardrailIdentifier="test-id",
guardrailVersion="DRAFT",
**extras,
)
def test_initialize_bedrock_wires_streaming_flags():
from litellm.proxy.guardrails.guardrail_initializers import initialize_bedrock
configured = initialize_bedrock(
_streaming_litellm_params(
streaming_buffer_until_moderated=False,
streaming_sampling_rate=3,
streaming_end_of_stream_only=True,
),
{"guardrail_name": "bedrock-streaming"},
)
defaulted = initialize_bedrock(
_streaming_litellm_params(),
{"guardrail_name": "bedrock-defaults"},
)
for registered in (configured, defaulted):
litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, registered)
assert configured.streaming_buffer_until_moderated is False
assert configured.streaming_sampling_rate == 3
assert configured.streaming_end_of_stream_only is True
assert defaulted.streaming_buffer_until_moderated is True
assert defaulted.streaming_sampling_rate == 5
assert defaulted.streaming_end_of_stream_only is False
def test_initialize_bedrock_rejects_non_positive_sampling_rate():
from pydantic import ValidationError
from litellm.proxy.guardrails.guardrail_initializers import initialize_bedrock
with pytest.raises(ValidationError):
initialize_bedrock(
_streaming_litellm_params(streaming_sampling_rate=0),
{"guardrail_name": "bedrock-bad-rate"},
)
def test_update_in_memory_litellm_params_round_trips_streaming_flags():
guardrail = BedrockGuardrail(
guardrail_name="bedrock-update",
guardrailIdentifier="test-id",
guardrailVersion="DRAFT",
)
guardrail.update_in_memory_litellm_params(
_streaming_litellm_params(
streaming_buffer_until_moderated=False,
streaming_sampling_rate=7,
streaming_end_of_stream_only=True,
)
)
assert guardrail.streaming_buffer_until_moderated is False
assert guardrail.streaming_sampling_rate == 7
assert guardrail.streaming_end_of_stream_only is True
guardrail.update_in_memory_litellm_params(_streaming_litellm_params())
assert guardrail.streaming_buffer_until_moderated is True
assert guardrail.streaming_sampling_rate == 5
assert guardrail.streaming_end_of_stream_only is False
async def _run_streaming_hook_recording_order(guardrail: BedrockGuardrail) -> list:
events = []
minimal = {"action": "NONE", "assessments": [], "outputs": []}
async def record_scan(*args, **kwargs):
events.append("scan")
return minimal
async def mock_stream():
yield _chat_chunk("Hello", None)
yield _chat_chunk(" world", None)
yield _chat_chunk("", "stop")
with patch.object(guardrail, "make_bedrock_api_request", AsyncMock(side_effect=record_scan)):
async for chunk in guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=UserAPIKeyAuth(),
response=mock_stream(),
request_data={"model": "gpt-4o-mini", "messages": [{"role": "user", "content": "hi"}]},
):
content = chunk.choices[0].delta.content if chunk.choices else None
events.append(("chunk", content))
return events
@pytest.mark.asyncio
async def test_unbuffered_end_of_stream_hook_yields_chunks_before_scan():
guardrail = BedrockGuardrail(
guardrail_name="bedrock-audit-mode",
guardrailIdentifier="test-id",
guardrailVersion="DRAFT",
event_hook=GuardrailEventHooks.post_call,
default_on=True,
streaming_buffer_until_moderated=False,
streaming_end_of_stream_only=True,
)
events = await _run_streaming_hook_recording_order(guardrail)
scan_index = events.index("scan")
chunk_events = [e for e in events if e != "scan"]
assert events.count("scan") == 1
assert [e for e in events[:scan_index] if e != "scan"] == chunk_events[: scan_index]
assert ("chunk", "Hello") in events[:scan_index]
assert ("chunk", " world") in events[:scan_index]
assert len(chunk_events) == 3
@pytest.mark.asyncio
async def test_buffered_default_hook_scans_before_any_chunk():
guardrail = BedrockGuardrail(
guardrail_name="bedrock-buffered-default",
guardrailIdentifier="test-id",
guardrailVersion="DRAFT",
event_hook=GuardrailEventHooks.post_call,
default_on=True,
)
events = await _run_streaming_hook_recording_order(guardrail)
assert events[0] == "scan"
assert all(e == "scan" or e[0] == "chunk" for e in events)
assert len([e for e in events if e != "scan"]) >= 1
@pytest.mark.asyncio
async def test_masking_keeps_buffered_path_even_when_unbuffered_configured():
guardrail = BedrockGuardrail(
guardrail_name="bedrock-mask-buffered",
guardrailIdentifier="test-id",
guardrailVersion="DRAFT",
event_hook=GuardrailEventHooks.post_call,
default_on=True,
mask_response_content=True,
streaming_buffer_until_moderated=False,
streaming_end_of_stream_only=True,
)
assert guardrail._streams_incrementally() is False
events = await _run_streaming_hook_recording_order(guardrail)
assert events[0] == "scan"
@pytest.mark.asyncio
async def test_streaming_end_of_stream_block_emits_error_frame_instead_of_truncating():
"""Regression for PR #38722: a topicPolicy DENY caught by the end-of-stream
scan used to raise after SSE headers were flushed, so the client saw a
silently truncated stream. The unified hook must emit the chat in-stream
error frame instead."""
from litellm.llms import load_guardrail_translation_mappings
from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail import (
unified_guardrail as unified_module,
)
from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import (
UnifiedLLMGuardrails,
)
from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
guardrail = BedrockGuardrail(
guardrailIdentifier="test-guardrail",
guardrailVersion="DRAFT",
streaming_end_of_stream_only=True,
streaming_buffer_until_moderated=False,
guardrail_name="bedrock-eos",
event_hook=GuardrailEventHooks.post_call,
default_on=True,
)
blocked_response = {
"action": "GUARDRAIL_INTERVENED",
"actionReason": "Guardrail blocked.",
"outputs": [{"text": "Sorry, the model cannot answer this question."}],
"assessments": [
{"topicPolicy": {"topics": [{"name": "Forbidden topic", "type": "DENY", "action": "BLOCKED"}]}}
],
}
def _chunk(content, finish_reason=None):
return ModelResponseStream(
choices=[
StreamingChoices(
index=0,
delta={"content": content, "role": "assistant"},
finish_reason=finish_reason,
)
],
)
async def _mock_stream():
yield _chunk("the forbidden ")
yield _chunk("topic answer", finish_reason="stop")
unified_module.endpoint_guardrail_translation_mappings = load_guardrail_translation_mappings()
try:
with patch.object(guardrail, "make_bedrock_api_request", new_callable=AsyncMock) as mock_api:
mock_api.side_effect = guardrail._get_http_exception_for_blocked_guardrail(blocked_response)
out = []
async for item in UnifiedLLMGuardrails().async_post_call_streaming_iterator_hook(
user_api_key_dict=UserAPIKeyAuth(api_key="test", request_route="/v1/chat/completions"),
response=_mock_stream(),
request_data={"guardrail_to_apply": guardrail, "model": "gpt-4"},
):
out.append(item)
finally:
unified_module.endpoint_guardrail_translation_mappings = None
assert len(out) == 3
assert isinstance(out[0], ModelResponseStream)
frame = out[-1]
assert isinstance(frame, bytes)
payload = json.loads(frame.decode()[len("data: ") :])
assert payload["error"]["message"] == "Violated guardrail policy"
assert payload["error"]["code"] == "400"
assert payload["error"]["provider_specific_fields"]["guardrailIdentifier"] == "test-guardrail"

7
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_hiddenlayer.py
View file

@ -432,7 +432,7 @@ class TestHiddenlayerGuardrail:
@pytest.mark.asyncio
async def test_apply_guardrail_request_with_image(self, monkeypatch: pytest.MonkeyPatch):
"""Test apply_guardrail sends multimodal content (image) to HiddenLayer v1."""
"""Test apply_guardrail strips images from multimodal content before sending to HiddenLayer v1."""
monkeypatch.setenv("HIDDENLAYER_API_BASE", "https://my.hiddenlayer")
guardrail = HiddenlayerGuardrail(
@ -485,12 +485,13 @@ class TestHiddenlayerGuardrail:
logging_obj=logging_obj,
)
# v1 API requires string content — multimodal list is stringified
# v1 API requires string content — image_url items are stripped and the
# remaining (text-only) content is stringified before being sent.
mock_post.assert_called_once()
call_kwargs = mock_post.call_args.kwargs
sent_content = call_kwargs["json"]["input"]["messages"][0]["content"]
assert isinstance(sent_content, str)
assert sent_content == str(multimodal_content)
assert sent_content == str([{"type": "text", "text": "how much is on this receipt?"}])
# Result should be returned without error
assert result is not None

235
tests/test_litellm/proxy/guardrails/guardrail_hooks/unified_guardrails/test_unified_guardrail.py
View file

@ -948,19 +948,24 @@ class TestStreamingTransform:
assert streamed == "ABCDEFGHIJ"
@pytest.mark.asyncio
async def test_incremental_diff_underflow_raises(self):
async def test_incremental_diff_underflow_emits_error_frame(self):
"""A transform shorter than what was already streamed cannot retract
bytes: it raises HTTPException(stream_transform_underflow)."""
bytes. Chunks have already been flushed by then, so the underflow
surfaces as the in-stream error frame, not an unraisable HTTPException."""
import json as _json
# First sample emits "ABCDEF" (6 chars); second sample shrinks to 3.
guardrail = _StreamingTextGuardrail(shrink_to="ABC", shrink_after=1)
chunks = [_stream_chunk("abcdef"), _stream_chunk("ghij")]
with pytest.raises(unified_module.HTTPException) as exc_info:
await _drive_stream(UnifiedLLMGuardrails(), guardrail, chunks)
out = await _drive_stream(UnifiedLLMGuardrails(), guardrail, chunks)
assert exc_info.value.status_code == 400
assert exc_info.value.detail["error"] == "stream_transform_underflow"
frame = out[-1]
assert isinstance(frame, bytes)
payload = _json.loads(frame.decode()[len("data: ") :])
assert payload["error"]["message"] == "stream_transform_underflow"
assert payload["error"]["code"] == "400"
@pytest.mark.asyncio
async def test_incremental_diff_final_chunk_preserves_finish_reason(self):
@ -1747,3 +1752,221 @@ class TestAppliedGuardrailsReflectsExecution:
async def test_ordinary_guardrail_is_auto_marked_applied(self):
data = await self._run(_AutoLoggingGuardrail())
assert "auto-logging" in _applied_guardrails(data)
class _EosHttpBlockingGuardrail(CustomGuardrail):
"""Raises the bedrock-shaped block HTTPException at end-of-stream scan time."""
def __init__(self):
super().__init__(guardrail_name="eos-http-block")
self.streaming_end_of_stream_only = True
def should_run_guardrail(self, data, event_type): # type: ignore[override]
return True
async def apply_guardrail(self, inputs, request_data, input_type, **kwargs):
raise unified_module.HTTPException(
status_code=400,
detail={
"error": "Violated guardrail policy",
"bedrock_guardrail_response": "BLOCKED_TOPIC",
},
)
def _anthropic_sse_event(event_type, data):
import json as _json
return f"event: {event_type}\ndata: {_json.dumps(data)}\n\n".encode()
def _anthropic_message_chunks(texts):
head = [
_anthropic_sse_event(
"message_start",
{
"type": "message_start",
"message": {
"id": "msg_test",
"type": "message",
"role": "assistant",
"model": "claude-sonnet-5",
"content": [],
"stop_reason": None,
"usage": {"input_tokens": 1, "output_tokens": 0},
},
},
),
_anthropic_sse_event(
"content_block_start",
{"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}},
),
]
deltas = [
_anthropic_sse_event(
"content_block_delta",
{"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": text}},
)
for text in texts
]
tail = [
_anthropic_sse_event("content_block_stop", {"type": "content_block_stop", "index": 0}),
_anthropic_sse_event(
"message_delta",
{
"type": "message_delta",
"delta": {"stop_reason": "end_turn", "stop_sequence": None},
"usage": {"output_tokens": 5},
},
),
_anthropic_sse_event("message_stop", {"type": "message_stop"}),
]
return head + deltas + tail
class TestStreamingHttpErrorFrames:
"""A post-flush end-of-stream guardrail block (HTTPException) must surface as
the endpoint's in-stream error frame instead of an unhandled raise that
silently truncates the SSE stream (PR #38722 defect 1)."""
@pytest.fixture(autouse=True)
def _use_real_mappings(self):
unified_module.endpoint_guardrail_translation_mappings = load_guardrail_translation_mappings()
yield
unified_module.endpoint_guardrail_translation_mappings = None
@pytest.mark.asyncio
async def test_chat_eos_block_emits_data_error_frame(self):
import json as _json
guardrail = _EosHttpBlockingGuardrail()
chunks = [_stream_chunk("hello "), _stream_chunk("world", finish_reason="stop")]
out = await _drive_stream(UnifiedLLMGuardrails(), guardrail, chunks)
assert out[:2] == chunks
frame = out[-1]
assert isinstance(frame, bytes)
text = frame.decode()
assert text.startswith("data: ")
payload = _json.loads(text[len("data: ") :])
assert payload["error"]["message"] == "Violated guardrail policy"
assert payload["error"]["code"] == "400"
@pytest.mark.asyncio
async def test_messages_eos_block_emits_anthropic_error_event(self):
guardrail = _EosHttpBlockingGuardrail()
chunks = _anthropic_message_chunks(["hello ", "world"])
out = await _drive_stream(
UnifiedLLMGuardrails(), guardrail, chunks, request_route="/v1/messages"
)
raw = b"".join(c for c in out if isinstance(c, bytes)).decode()
assert "hello " in raw
assert "event: error" in raw
assert "Violated guardrail policy" in raw
assert "guardrail_error" in raw
@pytest.mark.asyncio
async def test_responses_eos_block_emits_error_event_with_next_sequence(self):
guardrail = _EosHttpBlockingGuardrail()
chunks = [
{"type": "response.created", "sequence_number": 0},
{"type": "response.output_text.delta", "sequence_number": 1, "delta": "hello"},
{
"type": "response.completed",
"sequence_number": 2,
"response": {
"model": "gpt-4",
"output": [{"type": "message", "content": [{"type": "output_text", "text": "hello"}]}],
},
},
]
out = await _drive_stream(
UnifiedLLMGuardrails(), guardrail, chunks, request_route="/v1/responses"
)
assert chunks[0] in out and chunks[1] in out
assert chunks[2] not in out
error_event = out[-1]
assert error_event.type == "error"
assert error_event.sequence_number == 2
assert error_event.error.message == "Violated guardrail policy"
assert error_event.error.code == "400"
assert error_event.error.type == "guardrail_error"
@pytest.mark.asyncio
async def test_pre_flush_block_still_raises_http_exception(self):
guardrail = _EosHttpBlockingGuardrail()
guardrail.streaming_buffer_until_moderated = True
chunks = [_stream_chunk("hello "), _stream_chunk("world", finish_reason="stop")]
with pytest.raises(unified_module.HTTPException) as exc_info:
await _drive_stream(UnifiedLLMGuardrails(), guardrail, chunks)
assert exc_info.value.status_code == 400
assert exc_info.value.detail["error"] == "Violated guardrail policy"
class _AuditRecordingGuardrail(CustomGuardrail):
"""Successful scan that records guardrail_information, like a flags-on audit."""
def __init__(self):
super().__init__(guardrail_name="audit-recorder")
self.streaming_end_of_stream_only = True
def should_run_guardrail(self, data, event_type): # type: ignore[override]
return True
async def apply_guardrail(self, inputs, request_data, input_type, **kwargs):
self.add_standard_logging_guardrail_information_to_request_data(
guardrail_json_response={"action": "NONE"},
request_data=request_data,
guardrail_status="success",
)
return inputs
class TestStreamingGuardrailInformationBucket:
"""guardrail_information written during a chat streaming end-of-stream scan
must land in the request's ``metadata`` bucket that spend logging snapshots.
Regression for PR #38722 defect 2: the chat handler used to plant a
``litellm_metadata`` key first, flipping the bucket so every later
guardrail_information write was diverted and /spend/logs showed null."""
@pytest.fixture(autouse=True)
def _use_real_mappings(self):
unified_module.endpoint_guardrail_translation_mappings = load_guardrail_translation_mappings()
yield
unified_module.endpoint_guardrail_translation_mappings = None
@pytest.mark.asyncio
async def test_chat_eos_scan_writes_guardrail_information_to_metadata(self):
guardrail = _AuditRecordingGuardrail()
chunks = [_stream_chunk("hello "), _stream_chunk("world", finish_reason="stop")]
async def _mock_stream():
for chunk in chunks:
yield chunk
user_api_key_dict = UserAPIKeyAuth(
api_key="test-key", user_id="user-1", request_route="/v1/chat/completions"
)
request_data = {"guardrail_to_apply": guardrail, "model": "gpt-4", "metadata": {}}
out = []
async for item in UnifiedLLMGuardrails().async_post_call_streaming_iterator_hook(
user_api_key_dict=user_api_key_dict,
response=_mock_stream(),
request_data=request_data,
):
out.append(item)
assert "litellm_metadata" not in request_data
recorded = request_data["metadata"]["standard_logging_guardrail_information"]
assert len(recorded) == 1
assert recorded[0]["guardrail_name"] == "audit-recorder"
assert recorded[0]["guardrail_status"] == "success"
assert request_data["metadata"]["user_api_key_user_id"] == "user-1"

226
tests/test_litellm/proxy/guardrails/test_deferred_guardrail_logging.py
View file

@ -1228,3 +1228,229 @@ class TestFireDeferredStreamLogging:
assert info is not None, "guardrail_information should be populated"
assert len(info) == 1
assert info[0]["guardrail_name"] == "info-writer"
class TestResponsesIteratorDeferredLogging:
"""Regression for PR #38722 defect 2 on /v1/responses streams: when the
proxy arms _on_deferred_stream_complete, the responses streaming iterator
must store the logging coroutine for ProxyLogging._fire_deferred_stream_logging
(which runs AFTER end-of-stream guardrail scans write guardrail_information)
instead of dispatching immediately with a premature metadata snapshot."""
def _iterator(self, logging_obj):
from litellm.responses.streaming_iterator import (
BaseResponsesAPIStreamingIterator,
)
iterator = object.__new__(BaseResponsesAPIStreamingIterator)
iterator.logging_obj = logging_obj
iterator.start_time = None
iterator.completed_response = None
iterator._completed_response_logged = False
iterator._completed_response_cache_hit = None
iterator._persist_completed_response_before_logging = False
return iterator
def _logging_obj(self):
recorded = {}
async def dispatch_success_handlers(result=None, **kwargs):
recorded["dispatched"] = True
logging_obj = MagicMock()
logging_obj.dispatch_success_handlers = dispatch_success_handlers
return logging_obj, recorded
@pytest.mark.asyncio
async def test_armed_iterator_stores_deferred_coroutine(self):
logging_obj, recorded = self._logging_obj()
logging_obj._on_deferred_stream_complete = MagicMock()
iterator = self._iterator(logging_obj)
with patch("asyncio.create_task") as mock_create_task:
iterator._log_completed_response(is_async=True)
mock_create_task.assert_not_called()
args = logging_obj._deferred_stream_complete_args
assert isinstance(args, tuple) and len(args) == 1
assert "dispatched" not in recorded
await args[0]
assert recorded["dispatched"] is True
@pytest.mark.asyncio
async def test_unarmed_iterator_dispatches_immediately(self):
logging_obj, recorded = self._logging_obj()
logging_obj._on_deferred_stream_complete = None
iterator = self._iterator(logging_obj)
created = []
real_create_task = asyncio.create_task
def tracking_create_task(coro):
task = real_create_task(coro)
created.append(task)
return task
with patch("asyncio.create_task", side_effect=tracking_create_task):
iterator._log_completed_response(is_async=True)
assert len(created) == 1
await created[0]
assert recorded["dispatched"] is True
class TestArmDeferredStreamDispatch:
"""Regression for PR #38722: the closure shape armed on logging_obj must
match the args the stream's logging owner stores. Bridged /v1/responses
(LiteLLMCompletionStreamingIterator) shares its inner CustomStreamWrapper's
logging_obj, which stores (assembled_response, cache_hit); arming the
single-coroutine native closure there made _fire_deferred_stream_logging
raise TypeError inside the streaming hook, leaking an in-stream 500 error
frame on every streamed /v1/responses request."""
def _processor(self):
return ProxyBaseLLMRequestProcessing(data={"model": "gpt-test"})
def _dispatch_recording_logging_obj(self):
recorded = {}
async def dispatch_success_handlers(
result=None, start_time=None, end_time=None, cache_hit=None, prefer_async_handlers=False
):
recorded["result"] = result
recorded["cache_hit"] = cache_hit
recorded["prefer_async_handlers"] = prefer_async_handlers
logging_obj = MagicMock()
logging_obj.dispatch_success_handlers = dispatch_success_handlers
logging_obj._on_deferred_stream_complete = None
logging_obj._deferred_stream_complete_args = None
return logging_obj, recorded
@pytest.mark.asyncio
async def test_bridged_responses_iterator_gets_csw_arg_shape(self):
from litellm.responses.litellm_completion_transformation.streaming_iterator import (
LiteLLMCompletionStreamingIterator,
)
logging_obj, recorded = self._dispatch_recording_logging_obj()
bridged = object.__new__(LiteLLMCompletionStreamingIterator)
self._processor()._arm_deferred_stream_dispatch(
response=bridged,
route_type="aresponses",
user_api_key_dict=MagicMock(),
logging_obj=logging_obj,
)
assembled = object()
logging_obj._deferred_stream_complete_args = (assembled, False)
ProxyLogging._fire_deferred_stream_logging({"litellm_logging_obj": logging_obj})
await asyncio.sleep(0)
assert recorded["result"] is assembled
assert recorded["cache_hit"] is False
assert recorded["prefer_async_handlers"] is True
@pytest.mark.asyncio
async def test_router_wrapped_bridged_iterator_gets_csw_arg_shape(self):
"""The router wraps iterators without _hidden_params in
HiddenParamsAsyncIteratorWrapper before the proxy arms deferral, so
every production streamed /v1/responses reaches arming wrapped;
sniffing the wrapper instead of the inner iterator armed the 1-arg
native closure against the CSW's 2-arg stored shape and leaked a
TypeError 500 frame into the stream."""
from litellm.responses.litellm_completion_transformation.streaming_iterator import (
LiteLLMCompletionStreamingIterator,
)
from litellm.router_utils.add_retry_fallback_headers import (
HiddenParamsAsyncIteratorWrapper,
)
logging_obj, recorded = self._dispatch_recording_logging_obj()
wrapped = HiddenParamsAsyncIteratorWrapper(object.__new__(LiteLLMCompletionStreamingIterator))
self._processor()._arm_deferred_stream_dispatch(
response=wrapped,
route_type="aresponses",
user_api_key_dict=MagicMock(),
logging_obj=logging_obj,
)
assembled = object()
logging_obj._deferred_stream_complete_args = (assembled, False)
ProxyLogging._fire_deferred_stream_logging({"litellm_logging_obj": logging_obj})
await asyncio.sleep(0)
assert recorded["result"] is assembled
assert recorded["cache_hit"] is False
assert recorded["prefer_async_handlers"] is True
@pytest.mark.asyncio
async def test_native_stream_closure_enqueues_single_coroutine(self):
from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
logging_obj, _ = self._dispatch_recording_logging_obj()
async def _agen():
yield b"x"
self._processor()._arm_deferred_stream_dispatch(
response=_agen(),
route_type="anthropic_messages",
user_api_key_dict=MagicMock(),
logging_obj=logging_obj,
)
closure = logging_obj._on_deferred_stream_complete
assert closure is not None
async def _logging_coroutine():
return None
coro = _logging_coroutine()
with patch.object( # test-quality-ok: GLOBAL_LOGGING_WORKER is a process-global singleton with no injection seam
GLOBAL_LOGGING_WORKER, "ensure_initialized_and_enqueue"
) as mock_enqueue:
await closure(coro)
mock_enqueue.assert_called_once_with(async_coroutine=coro)
coro.close()
@pytest.mark.asyncio
async def test_csw_closure_routes_through_deferred_stream_guardrails(self, monkeypatch):
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
logging_obj, recorded = self._dispatch_recording_logging_obj()
csw = object.__new__(CustomStreamWrapper)
processor = self._processor()
monkeypatch.setattr( # test-quality-ok: empty the process-global callback registry so no ambient guardrail runs
litellm, "callbacks", []
)
processor._arm_deferred_stream_dispatch(
response=csw,
route_type="acompletion",
user_api_key_dict=MagicMock(),
logging_obj=logging_obj,
)
assembled = object()
await logging_obj._on_deferred_stream_complete(assembled, False)
await asyncio.sleep(0)
assert recorded["result"] is assembled
assert recorded["cache_hit"] is False
assert recorded["prefer_async_handlers"] is True
def test_non_native_route_generator_not_armed(self):
logging_obj, _ = self._dispatch_recording_logging_obj()
async def _agen():
yield b"x"
self._processor()._arm_deferred_stream_dispatch(
response=_agen(),
route_type="acompletion",
user_api_key_dict=MagicMock(),
logging_obj=logging_obj,
)
assert logging_obj._on_deferred_stream_complete is None

103
tests/test_litellm/proxy/guardrails/test_prompt_security_guardrails.py
View file

@ -1,16 +1,16 @@
from fastapi.exceptions import HTTPException
from unittest.mock import patch, AsyncMock
from httpx import Response, Request
import asyncio
import base64
from unittest.mock import AsyncMock, patch
import pytest
from litellm.proxy.guardrails.guardrail_hooks.prompt_security.prompt_security import (
PromptSecurityGuardrailMissingSecrets,
PromptSecurityGuardrail,
)
from fastapi.exceptions import HTTPException
from httpx import ReadTimeout, Request, Response
import litellm
from litellm.proxy.guardrails.guardrail_hooks.prompt_security.prompt_security import (
PromptSecurityGuardrail,
PromptSecurityGuardrailMissingSecrets,
)
from litellm.proxy.guardrails.init_guardrails import init_guardrails_v2
@ -30,6 +30,7 @@ def test_prompt_security_guard_config(monkeypatch: pytest.MonkeyPatch):
"guardrail": "prompt_security",
"mode": "during_call",
"default_on": True,
"file_sanitization_fail_open": False,
},
}
],
@ -41,6 +42,10 @@ def test_prompt_security_guard_config(monkeypatch: pytest.MonkeyPatch):
assert registered[0].guardrail_name == "prompt_security"
assert registered[0].default_on is True
assert registered[0].event_hook == "during_call"
assert registered[0].file_sanitization_fail_open is False
config_model = registered[0].get_config_model()
assert config_model is not None
assert config_model().file_sanitization_fail_open is True
def test_prompt_security_guard_config_no_api_key(monkeypatch: pytest.MonkeyPatch):
@ -374,6 +379,86 @@ async def test_file_sanitization(monkeypatch: pytest.MonkeyPatch):
assert result is not None
@pytest.mark.asyncio
@pytest.mark.parametrize(
"timeout",
(
litellm.Timeout(
message="Prompt Security upload timed out",
model="default-model-name",
llm_provider="litellm-httpx-handler",
),
ReadTimeout(
"Prompt Security poll timed out",
request=Request(method="GET", url="https://test.prompt.security/api/sanitizeFile"),
),
),
ids=("litellm", "httpx"),
)
@pytest.mark.parametrize("fail_open", (True, False), ids=("fail-open", "fail-closed"))
async def test_file_sanitization_request_timeout_policy(
monkeypatch: pytest.MonkeyPatch, timeout: Exception, fail_open: bool
):
monkeypatch.setenv("PROMPT_SECURITY_API_KEY", "test-key")
monkeypatch.setenv("PROMPT_SECURITY_API_BASE", "https://test.prompt.security")
guardrail = PromptSecurityGuardrail(
guardrail_name="test-guard",
event_hook="pre_call",
default_on=True,
file_sanitization_fail_open=fail_open,
)
with patch.object(guardrail.async_handler, "post", AsyncMock(side_effect=timeout)):
if not fail_open:
with pytest.raises(HTTPException) as exc_info:
await guardrail.sanitize_file_content(b"file-content", "document.pdf")
assert exc_info.value.status_code == 408
assert exc_info.value.detail == "File sanitization timeout"
return
result = await guardrail.sanitize_file_content(b"file-content", "document.pdf")
assert result == {
"action": "allow",
"content": None,
"metadata": {},
"violations": (),
}
@pytest.mark.asyncio
@pytest.mark.parametrize("fail_open", (True, False), ids=("fail-open", "fail-closed"))
async def test_file_sanitization_overall_timeout_policy(monkeypatch: pytest.MonkeyPatch, fail_open: bool):
monkeypatch.setenv("PROMPT_SECURITY_API_KEY", "test-key")
monkeypatch.setenv("PROMPT_SECURITY_API_BASE", "https://test.prompt.security")
guardrail = PromptSecurityGuardrail(
guardrail_name="test-guard",
event_hook="pre_call",
default_on=True,
file_sanitization_timeout=0.01,
file_sanitization_fail_open=fail_open,
)
async def hanging_post(*_args: object, **_kwargs: object) -> None:
await asyncio.sleep(60)
raise AssertionError("sanitization request should have been cancelled")
with patch.object(guardrail.async_handler, "post", side_effect=hanging_post):
if not fail_open:
with pytest.raises(HTTPException) as exc_info:
await guardrail.sanitize_file_content(b"file-content", "document.pdf")
assert exc_info.value.status_code == 408
assert exc_info.value.detail == "File sanitization timeout"
return
result = await guardrail.sanitize_file_content(b"file-content", "document.pdf")
assert result["action"] == "allow"
assert result["content"] is None
@pytest.mark.asyncio
async def test_file_sanitization_block(monkeypatch: pytest.MonkeyPatch):
"""Test that file sanitization blocks malicious files"""
@ -544,7 +629,7 @@ async def test_role_filtering(monkeypatch: pytest.MonkeyPatch):
return mock_response
with patch.object(guardrail.async_handler, "post", side_effect=mock_post):
result = await guardrail.apply_guardrail(
await guardrail.apply_guardrail(
inputs=inputs,
request_data=request_data,
input_type="request",

44
tests/test_litellm/proxy/management_endpoints/test_customer_endpoints.py
View file

@ -83,6 +83,50 @@ def test_update_customer_success(mock_prisma_client, mock_user_api_key_auth):
assert response.json()["alias"] == "Updated Test User"
def test_update_customer_unblock(mock_prisma_client, mock_user_api_key_auth):
mock_end_user = LiteLLM_EndUserTable(user_id="test-user-1", blocked=True)
updated_mock_end_user = LiteLLM_EndUserTable(user_id="test-user-1", blocked=False)
mock_prisma_client.db.litellm_endusertable.find_first = AsyncMock(return_value=mock_end_user)
mock_prisma_client.db.litellm_endusertable.update = AsyncMock(return_value=updated_mock_end_user)
response = client.post(
"/customer/update",
json={"user_id": "test-user-1", "blocked": False},
headers={"Authorization": "Bearer test-key"},
)
assert response.status_code == 200
assert response.json()["blocked"] is False
update_mock = mock_prisma_client.db.litellm_endusertable.update
update_mock.assert_called_once()
assert update_mock.call_args.kwargs["data"]["blocked"] is False
def test_update_customer_keeps_blocked_when_omitted(mock_prisma_client, mock_user_api_key_auth):
"""
Regression test: updating a blocked customer without supplying `blocked`
must NOT reset it to unblocked. `blocked=False` is the model default and
should only be applied when explicitly provided by the caller.
"""
mock_end_user = LiteLLM_EndUserTable(user_id="test-user-1", blocked=True)
updated_mock_end_user = LiteLLM_EndUserTable(user_id="test-user-1", blocked=True)
mock_prisma_client.db.litellm_endusertable.find_first = AsyncMock(return_value=mock_end_user)
mock_prisma_client.db.litellm_endusertable.update = AsyncMock(return_value=updated_mock_end_user)
response = client.post(
"/customer/update",
json={"user_id": "test-user-1", "alias": "Updated Test User"},
headers={"Authorization": "Bearer test-key"},
)
assert response.status_code == 200
update_mock = mock_prisma_client.db.litellm_endusertable.update
update_mock.assert_called_once()
assert "blocked" not in update_mock.call_args.kwargs["data"]
def test_update_customer_not_found(mock_prisma_client, mock_user_api_key_auth):
"""
Test that update_end_user raises a 404 ProxyException when user_id does not exist.

133
tests/test_litellm/proxy/pass_through_endpoints/test_streaming_handler_interrupt.py
View file

@ -28,12 +28,21 @@ def _make_streaming_response(chunks):
return mock
def _unarmed_logging_obj():
"""Real Logging objects only carry _on_deferred_stream_complete when the
proxy arms deferred dispatch; a bare MagicMock's auto-attribute is truthy
and would spuriously trigger the deferral branch."""
obj = MagicMock()
obj._on_deferred_stream_complete = None
return obj
@pytest.mark.asyncio
async def test_chunk_processor_logs_on_normal_completion():
chunks = [b"chunk-1", b"chunk-2", b"chunk-3"]
response = _make_streaming_response(chunks)
mock_logging_obj = MagicMock()
mock_logging_obj = _unarmed_logging_obj()
mock_passthrough_handler = MagicMock()
with patch.object(
@ -66,7 +75,7 @@ async def test_chunk_processor_logs_on_client_disconnect():
chunks = [b"event-1", b"event-2", b"event-3"]
response = _make_streaming_response(chunks)
mock_logging_obj = MagicMock()
mock_logging_obj = _unarmed_logging_obj()
mock_passthrough_handler = MagicMock()
with patch.object(
@ -104,7 +113,7 @@ async def test_chunk_processor_does_not_schedule_success_logging_for_upstream_er
response = _make_streaming_response(chunks)
response.status_code = 403
mock_logging_obj = MagicMock()
mock_logging_obj = _unarmed_logging_obj()
mock_passthrough_handler = MagicMock()
with patch.object(
@ -134,7 +143,7 @@ async def test_chunk_processor_does_not_schedule_success_logging_for_upstream_er
async def test_chunk_processor_does_not_schedule_logging_when_no_chunks():
response = _make_streaming_response([])
mock_logging_obj = MagicMock()
mock_logging_obj = _unarmed_logging_obj()
mock_passthrough_handler = MagicMock()
with patch.object(
@ -189,7 +198,7 @@ async def test_chunk_processor_routes_logging_through_logging_worker():
async for chunk in PassThroughStreamingHandler.chunk_processor(
response=response,
request_body={"model": "claude-3-haiku"},
litellm_logging_obj=MagicMock(),
litellm_logging_obj=_unarmed_logging_obj(),
endpoint_type=EndpointType.GENERIC,
start_time=datetime.now(),
passthrough_success_handler_obj=MagicMock(),
@ -230,7 +239,7 @@ async def test_chunk_processor_routes_logging_through_logging_worker_on_disconne
gen = PassThroughStreamingHandler.chunk_processor(
response=response,
request_body={"model": "claude-3-haiku"},
litellm_logging_obj=MagicMock(),
litellm_logging_obj=_unarmed_logging_obj(),
endpoint_type=EndpointType.GENERIC,
start_time=datetime.now(),
passthrough_success_handler_obj=MagicMock(),
@ -246,7 +255,7 @@ async def test_chunk_processor_routes_logging_through_logging_worker_on_disconne
def _logging_obj_with_write_once_cst():
"""Build a MagicMock that mirrors the real Logging behavior: _update_completion_start_time
latches self.completion_start_time so the write-once guard actually latches."""
obj = MagicMock()
obj = _unarmed_logging_obj()
obj.completion_start_time = None
def _update(*, completion_start_time):
@ -301,7 +310,7 @@ async def test_chunk_processor_does_not_reset_completion_start_time_on_later_chu
response = _make_streaming_response(chunks)
real_first = datetime(2020, 1, 1, 0, 0, 0)
mock_logging_obj = MagicMock()
mock_logging_obj = _unarmed_logging_obj()
# Simulate first-chunk stamp having already landed (e.g. under contention or a
# prior wrapper that already set it): later chunks must be no-ops.
mock_logging_obj.completion_start_time = real_first
@ -387,7 +396,7 @@ async def _collect_openai_passthrough_chunks(chunks, endpoint_type):
async for chunk in PassThroughStreamingHandler.chunk_processor(
response=response,
request_body={"model": "gpt-4o-mini", "stream": True},
litellm_logging_obj=MagicMock(),
litellm_logging_obj=_unarmed_logging_obj(),
endpoint_type=endpoint_type,
start_time=datetime.now(),
passthrough_success_handler_obj=MagicMock(),
@ -517,3 +526,109 @@ def test_convert_raw_bytes_survives_truncated_multibyte_sequence():
lines = PassThroughStreamingHandler._convert_raw_bytes_to_str_lines(raw_bytes)
assert any('"type": "message_delta"' in line for line in lines)
@pytest.mark.asyncio
async def test_chunk_processor_defers_logging_until_fire_when_armed():
"""Regression for PR #38722: native /v1/messages streams route through
chunk_processor, which enqueued the spend log the moment the stream ended,
racing the guardrail end-of-stream scan and logging
guardrail_information as null. With deferred dispatch armed, the completed
stream must park the logging coroutine on logging_obj and only enqueue it
when ProxyLogging._fire_deferred_stream_logging fires after the scan."""
from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
from litellm.proxy.utils import ProxyLogging
chunks = [b"event-1", b"event-2"]
response = _make_streaming_response(chunks)
logging_obj = _unarmed_logging_obj()
logging_obj._deferred_stream_complete_args = None
enqueued = []
def _capture(async_coroutine):
enqueued.append(async_coroutine)
async_coroutine.close()
with patch.object( # test-quality-ok: GLOBAL_LOGGING_WORKER is a process-global singleton with no injection seam
GLOBAL_LOGGING_WORKER,
"ensure_initialized_and_enqueue",
side_effect=_capture,
) as mock_enqueue:
gen = PassThroughStreamingHandler.chunk_processor(
response=response,
request_body={"model": "claude-3-haiku"},
litellm_logging_obj=logging_obj,
endpoint_type=EndpointType.ANTHROPIC,
start_time=datetime.now(),
passthrough_success_handler_obj=MagicMock(),
url_route="/v1/messages",
route_streaming_logging=AsyncMock(),
)
ProxyBaseLLMRequestProcessing(data={})._arm_deferred_stream_dispatch(
response=gen,
route_type="anthropic_messages",
user_api_key_dict=MagicMock(),
logging_obj=logging_obj,
)
received = []
async for chunk in gen:
received.append(chunk)
await asyncio.sleep(0)
assert received == chunks
mock_enqueue.assert_not_called()
parked = logging_obj._deferred_stream_complete_args
assert isinstance(parked, tuple) and len(parked) == 1
assert asyncio.iscoroutine(parked[0])
ProxyLogging._fire_deferred_stream_logging({"litellm_logging_obj": logging_obj})
await asyncio.sleep(0)
mock_enqueue.assert_called_once()
@pytest.mark.asyncio
async def test_chunk_processor_enqueues_immediately_on_disconnect_even_when_armed():
"""Client disconnects never reach _fire_deferred_stream_logging, so parking
the coroutine there would lose the partial-usage spend log (LIT-2642); the
disconnect path must keep enqueueing immediately."""
chunks = [b"event-1", b"event-2", b"event-3"]
response = _make_streaming_response(chunks)
logging_obj = _unarmed_logging_obj()
async def _armed_closure(logging_coroutine):
raise AssertionError("deferred closure must not fire on disconnect")
logging_obj._on_deferred_stream_complete = _armed_closure
logging_obj._deferred_stream_complete_args = None
enqueued = []
def _capture(async_coroutine):
enqueued.append(async_coroutine)
async_coroutine.close()
with patch.object( # test-quality-ok: GLOBAL_LOGGING_WORKER is a process-global singleton with no injection seam
GLOBAL_LOGGING_WORKER,
"ensure_initialized_and_enqueue",
side_effect=_capture,
) as mock_enqueue:
gen = PassThroughStreamingHandler.chunk_processor(
response=response,
request_body={"model": "claude-3-haiku"},
litellm_logging_obj=logging_obj,
endpoint_type=EndpointType.ANTHROPIC,
start_time=datetime.now(),
passthrough_success_handler_obj=MagicMock(),
url_route="/v1/messages",
route_streaming_logging=AsyncMock(),
)
await gen.__anext__()
await gen.aclose()
mock_enqueue.assert_called_once()
assert logging_obj._deferred_stream_complete_args is None

237
tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
View file

@ -3,10 +3,12 @@ Test for response_api_endpoints/endpoints.py
"""
import unittest
from typing import Any
from unittest.mock import AsyncMock, MagicMock, patch
import pytest
from fastapi.testclient import TestClient
from httpx import Response
import litellm
from litellm.proxy.proxy_server import app
@ -82,11 +84,7 @@ class TestResponsesAPIEndpoints(unittest.TestCase):
ResponseOutputMessage(
type="message",
role="assistant",
content=[
ResponseOutputText(
type="output_text", text="Hello from Cursor!"
)
],
content=[ResponseOutputText(type="output_text", text="Hello from Cursor!")],
)
],
)
@ -121,9 +119,7 @@ class TestResponsesAPIEndpoints(unittest.TestCase):
@pytest.mark.asyncio
@patch("litellm.proxy.proxy_server.llm_router")
@patch("litellm.proxy.proxy_server.user_api_key_auth")
async def test_responses_api_key_spend_header_includes_response_cost(
self, mock_auth, mock_router
):
async def test_responses_api_key_spend_header_includes_response_cost(self, mock_auth, mock_router):
"""
Test that x-litellm-key-spend header includes the current request's response_cost
for /v1/responses endpoint.
@ -159,9 +155,7 @@ class TestResponsesAPIEndpoints(unittest.TestCase):
ResponseOutputMessage(
type="message",
role="assistant",
content=[
ResponseOutputText(type="output_text", text="Test response")
],
content=[ResponseOutputText(type="output_text", text="Test response")],
)
],
)
@ -356,6 +350,7 @@ class TestWSModelExtraction:
from litellm.proxy.response_api_endpoints.endpoints import (
_extract_model_from_first_ws_event,
)
event = {"type": "response.create", "model": "gpt-4o", "input": "hello"}
assert _extract_model_from_first_ws_event(event) == "gpt-4o"
@ -363,6 +358,7 @@ class TestWSModelExtraction:
from litellm.proxy.response_api_endpoints.endpoints import (
_extract_model_from_first_ws_event,
)
event = {"type": "response.create", "response": {"model": "gpt-4o", "input": "hello"}}
assert _extract_model_from_first_ws_event(event) == "gpt-4o"
@ -370,6 +366,7 @@ class TestWSModelExtraction:
from litellm.proxy.response_api_endpoints.endpoints import (
_extract_model_from_first_ws_event,
)
event = {
"type": "response.create",
"model": "flat-model",
@ -381,6 +378,7 @@ class TestWSModelExtraction:
from litellm.proxy.response_api_endpoints.endpoints import (
_extract_model_from_first_ws_event,
)
event = {"type": "response.create", "input": "hello"}
assert _extract_model_from_first_ws_event(event) is None
@ -400,9 +398,7 @@ class TestResponsesWSFirstFrameValidation:
)
ws = MagicMock()
ws.receive_text = AsyncMock(
return_value=json.dumps({"type": "session.update", "model": "gpt-4o"})
)
ws.receive_text = AsyncMock(return_value=json.dumps({"type": "session.update", "model": "gpt-4o"}))
ws.send_text = AsyncMock()
ws.close = AsyncMock()
@ -412,10 +408,7 @@ class TestResponsesWSFirstFrameValidation:
ws.send_text.assert_awaited_once()
ws.close.assert_awaited_once_with(code=1008, reason="Invalid first message")
error_payload = json.loads(ws.send_text.await_args.args[0])
assert (
error_payload["error"]["message"]
== "First message must be a response.create JSON object."
)
assert error_payload["error"]["message"] == "First message must be a response.create JSON object."
@pytest.mark.asyncio
async def test_rejects_non_object_json_first_frame(self):
@ -484,16 +477,12 @@ class TestResponsesWSFirstFrameModelAuth:
ws.url = "ws://testserver/v1/responses"
ws.accept = AsyncMock()
ws.receive_text = AsyncMock(
return_value=json.dumps(
{"type": "response.create", "model": "gpt-4o-mini", "input": []}
)
return_value=json.dumps({"type": "response.create", "model": "gpt-4o-mini", "input": []})
)
ws.close = AsyncMock()
processor = MagicMock()
processor.common_processing_pre_call_logic = AsyncMock(
return_value=({"model": "gpt-4o-mini"}, MagicMock())
)
processor.common_processing_pre_call_logic = AsyncMock(return_value=({"model": "gpt-4o-mini"}, MagicMock()))
async def fake_llm_call():
return None
@ -529,9 +518,7 @@ class TestResponsesWSFirstFrameModelAuth:
_enforce_responses_ws_first_frame_model_auth,
)
request = Request(
{"type": "http", "method": "POST", "path": "/v1/responses", "headers": []}
)
request = Request({"type": "http", "method": "POST", "path": "/v1/responses", "headers": []})
user_api_key_dict = MagicMock()
llm_router = MagicMock()
@ -593,9 +580,7 @@ class TestReadWSModelFromFirstFrameErrors:
assert result is None
ws.send_text.assert_not_awaited()
ws.close.assert_awaited_once_with(
code=1008, reason="Timed out waiting for first message"
)
ws.close.assert_awaited_once_with(code=1008, reason="Timed out waiting for first message")
@pytest.mark.asyncio
async def test_invalid_json_sends_error_and_closes(self):
@ -613,9 +598,7 @@ class TestReadWSModelFromFirstFrameErrors:
assert result is None
payload = json.loads(ws.send_text.await_args.args[0])
assert payload["error"]["message"] == "First message is not valid JSON."
ws.close.assert_awaited_once_with(
code=1008, reason="Invalid JSON in first message"
)
ws.close.assert_awaited_once_with(code=1008, reason="Invalid JSON in first message")
@pytest.mark.asyncio
async def test_missing_model_sends_error_and_closes(self):
@ -624,9 +607,7 @@ class TestReadWSModelFromFirstFrameErrors:
)
ws = MagicMock()
ws.receive_text = AsyncMock(
return_value=json.dumps({"type": "response.create", "input": []})
)
ws.receive_text = AsyncMock(return_value=json.dumps({"type": "response.create", "input": []}))
ws.send_text = AsyncMock()
ws.close = AsyncMock()
@ -679,10 +660,7 @@ class TestManagedResponsesSameProvider:
assert self._handler("gpt-4o")._same_provider("gpt-4o-mini") is True
def test_different_provider_is_not_same(self):
assert (
self._handler("gpt-4o")._same_provider("vertex_ai/gemini-2.0-flash")
is False
)
assert self._handler("gpt-4o")._same_provider("vertex_ai/gemini-2.0-flash") is False
def test_inject_credentials_keeps_provider_for_same_provider_model(self):
handler = self._handler("gpt-4o", custom_llm_provider="openai")
@ -697,18 +675,14 @@ class TestManagedResponsesSameProvider:
assert "custom_llm_provider" not in call_kwargs
def test_unresolvable_connection_model_falls_back_to_custom_provider(self):
handler = self._handler(
"my-custom-deployment", custom_llm_provider="openai"
)
handler = self._handler("my-custom-deployment", custom_llm_provider="openai")
assert handler._same_provider("gpt-4o-mini") is True
call_kwargs: dict = {}
handler._inject_credentials(call_kwargs, model="gpt-4o-mini")
assert call_kwargs["custom_llm_provider"] == "openai"
def test_unresolvable_connection_model_still_drops_cross_provider(self):
handler = self._handler(
"my-custom-deployment", custom_llm_provider="openai"
)
handler = self._handler("my-custom-deployment", custom_llm_provider="openai")
call_kwargs: dict = {}
handler._inject_credentials(call_kwargs, model="vertex_ai/gemini-2.0-flash")
assert "custom_llm_provider" not in call_kwargs
@ -840,9 +814,7 @@ def test_cursor_chat_completions_input_body_uses_responses_pipeline_and_strips_s
type="message",
role="assistant",
status="completed",
content=[
ResponseOutputText(type="output_text", text="agent reply", annotations=[])
],
content=[ResponseOutputText(type="output_text", text="agent reply", annotations=[])],
)
],
)
@ -851,9 +823,12 @@ def test_cursor_chat_completions_input_body_uses_responses_pipeline_and_strips_s
app.dependency_overrides[user_api_key_auth] = _auth_override
try:
with patch.object(ps, "llm_router", mock_router), patch(
"litellm.proxy.response_api_endpoints.endpoints._read_request_body",
side_effect=capturing_read_request_body,
with (
patch.object(ps, "llm_router", mock_router),
patch(
"litellm.proxy.response_api_endpoints.endpoints._read_request_body",
side_effect=capturing_read_request_body,
),
):
client = TestClient(app)
response = client.post(
@ -1488,8 +1463,8 @@ def _router_serving_only(base_model: str) -> MagicMock:
mock_router.router_general_settings.pass_through_all_models = False
mock_router.default_deployment = None
mock_router.pattern_router.patterns = {base_model: ["anthropic/*"]}
mock_router.pattern_router.get_pattern.side_effect = (
lambda model: [{"model_name": "anthropic/*"}] if model == base_model else None
mock_router.pattern_router.get_pattern.side_effect = lambda model: (
[{"model_name": "anthropic/*"}] if model == base_model else None
)
return mock_router
@ -1739,9 +1714,7 @@ class TestCursorGateRecognizesRoutingGroups:
from litellm.proxy.response_api_endpoints.endpoints import _resolve_cursor_model_variant
router = Router(
model_list=[
{"model_name": "member-fast", "litellm_params": {"model": "openai/gpt-4o", "api_key": "fake"}}
],
model_list=[{"model_name": "member-fast", "litellm_params": {"model": "openai/gpt-4o", "api_key": "fake"}}],
routing_groups=[
{"group_name": "grouped-thinking-high", "models": ["member-fast"], "routing_strategy": "simple-shuffle"}
],
@ -1836,3 +1809,153 @@ class TestGuardrailBlockedResponsesUsage:
assert usage["input_tokens"] == 0
assert usage["output_tokens"] == 0
assert usage["total_tokens"] == 0
class TestResponsesInputTokens:
"""Regression tests for POST /v1/responses/input_tokens.
The docs promise OpenAI-format token counting on the proxy, but the route was
never registered, so the POST fell through to the GET/DELETE-only
/v1/responses/{response_id} route and returned 405."""
def _post_input_tokens(
self,
body: dict[str, Any],
path: str = "/v1/responses/input_tokens",
counter: AsyncMock | None = None,
) -> tuple[Response, AsyncMock]:
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
from litellm.proxy.response_api_endpoints.endpoints import _proxy_token_counter
from litellm.types.utils import TokenCountResponse
token_counter_mock = (
counter
if counter is not None
else AsyncMock(
return_value=TokenCountResponse(
total_tokens=13,
request_model=body.get("model", ""),
model_used=body.get("model", ""),
tokenizer_type="openai_api",
)
)
)
app.dependency_overrides[user_api_key_auth] = lambda: UserAPIKeyAuth(api_key="sk-test", request_route=path)
app.dependency_overrides[_proxy_token_counter] = lambda: token_counter_mock
try:
client = TestClient(app)
response = client.post(path, json=body, headers={"Authorization": "Bearer sk-1234"})
return response, token_counter_mock
finally:
app.dependency_overrides.pop(user_api_key_auth, None)
app.dependency_overrides.pop(_proxy_token_counter, None)
def test_string_input_returns_openai_input_tokens_shape(self):
response, counter = self._post_input_tokens({"model": "gpt-4o", "input": "Hello, how are you?"})
assert response.status_code == 200, response.text
assert response.json() == {"object": "response.input_tokens", "input_tokens": 13}
counter.assert_awaited_once()
assert counter.call_args.kwargs["call_endpoint"] is True
token_request = counter.call_args.kwargs["request"]
assert token_request.model == "gpt-4o"
assert token_request.messages == [{"role": "user", "content": "Hello, how are you?"}]
def test_every_route_alias_is_registered(self):
for path in ("/v1/responses/input_tokens", "/responses/input_tokens", "/openai/v1/responses/input_tokens"):
response, _ = self._post_input_tokens({"model": "gpt-4o", "input": "hi"}, path=path)
assert response.status_code == 200, f"{path}: {response.status_code} {response.text}"
def test_input_items_instructions_and_tools_are_forwarded(self):
tools = [
{
"type": "function",
"name": "get_weather",
"description": "Get weather for a city",
"parameters": {"type": "object", "properties": {"city": {"type": "string"}}},
}
]
response, counter = self._post_input_tokens(
{
"model": "gpt-4o",
"input": [{"role": "user", "content": "What is the weather in Paris?"}],
"instructions": "You are terse.",
"tools": tools,
}
)
assert response.status_code == 200, response.text
token_request = counter.call_args.kwargs["request"]
assert token_request.messages == [
{"role": "system", "content": "You are terse."},
{"role": "user", "content": "What is the weather in Paris?"},
]
assert token_request.tools == tools
def test_missing_model_returns_openai_400(self):
response, counter = self._post_input_tokens({"input": "Hello"})
assert response.status_code == 400, response.text
assert response.json() == {
"error": {
"message": "Missing required parameter: 'model'.",
"type": "invalid_request_error",
"param": "model",
"code": "missing_required_parameter",
}
}
counter.assert_not_awaited()
def test_missing_input_returns_openai_400(self):
response, counter = self._post_input_tokens({"model": "gpt-4o"})
assert response.status_code == 400, response.text
assert response.json() == {
"error": {
"message": "Missing required parameter: 'input'.",
"type": "invalid_request_error",
"param": "input",
"code": "missing_required_parameter",
}
}
counter.assert_not_awaited()
@pytest.mark.parametrize("empty_input", ["", []])
def test_empty_input_returns_openai_400(self, empty_input):
response, counter = self._post_input_tokens({"model": "gpt-4o", "input": empty_input})
assert response.status_code == 400, response.text
assert response.json() == {
"error": {
"message": """One of "input" or "previous_response_id" or 'prompt' or 'conversation' must be provided.""",
"type": "invalid_request_error",
"param": None,
"code": "missing_required_parameter",
}
}
counter.assert_not_awaited()
def test_invalid_tools_returns_openai_400(self):
response, counter = self._post_input_tokens({"model": "gpt-4o", "input": "hi", "tools": "not-a-list"})
assert response.status_code == 400, response.text
error = response.json()["error"]
assert error["type"] == "invalid_request_error"
counter.assert_not_awaited()
def test_provider_error_maps_status_code(self):
from litellm.proxy._types import ProxyException
failing_counter = AsyncMock(
side_effect=ProxyException(
message="rate limited",
type="token_counting_error",
param="model",
code="429",
)
)
response, _ = self._post_input_tokens({"model": "gpt-4o", "input": "hi"}, counter=failing_counter)
assert response.status_code == 429, response.text
assert response.json()["error"]["message"] == "rate limited"

48
tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py Normal file
View file

@ -0,0 +1,48 @@
from typing import Final
import pytest
from litellm.caching import DualCache
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
from litellm.proxy.spend_tracking.budget_reservation import reserve_budget_for_request
from litellm.proxy.utils import ProxyLogging
TOKEN_COUNTING_ROUTES: Final = (
"/responses/input_tokens",
"/v1/responses/input_tokens",
"/openai/v1/responses/input_tokens",
"/utils/token_counter",
)
def _budgeted_token() -> UserAPIKeyAuth:
return UserAPIKeyAuth(api_key="sk-test", token="hashed-token", max_budget=100.0, spend=0.0)
async def _reserve(route: str) -> dict | None:
return await reserve_budget_for_request(
request_body={"model": "gpt-4o", "input": "hello"},
route=route,
llm_router=None,
valid_token=_budgeted_token(),
team_object=None,
user_object=None,
prisma_client=None,
user_api_key_cache=UserApiKeyCache(),
proxy_logging_obj=ProxyLogging(user_api_key_cache=DualCache()),
)
@pytest.mark.asyncio
@pytest.mark.parametrize("route", TOKEN_COUNTING_ROUTES)
async def test_token_counting_routes_are_exempt_from_budget_reservation(route):
assert await _reserve(route) is None
@pytest.mark.asyncio
async def test_non_exempt_llm_route_still_reserves_budget():
reservation: Final = await _reserve("/v1/responses")
assert reservation is not None
assert reservation["reserved_cost"] > 0

6
tests/test_litellm/proxy/spend_tracking/test_spend_management_endpoints.py
View file

@ -2865,7 +2865,7 @@ class TestSpendLogsPayload:
"model": "gpt-4o",
"user": "",
"team_id": "",
"metadata": '{"applied_guardrails": [], "attempted_fallbacks": null, "original_model_group": null, "batch_models": null, "batch_successful_requests": null, "batch_failed_requests": null, "mcp_tool_call_metadata": null, "vector_store_request_metadata": null, "routing_decision": null, "internal_call_origin": null, "guardrail_information": null, "compression_savings": null, "litellm_gateway_injected_cache": null, "usage_object": {"completion_tokens": 20, "prompt_tokens": 10, "total_tokens": 30, "completion_tokens_details": null, "prompt_tokens_details": null}, "model_map_information": {"model_map_key": "gpt-4o", "model_map_value": {"key": "gpt-4o", "max_tokens": 16384, "max_input_tokens": 128000, "max_output_tokens": 16384, "input_cost_per_token": 2.5e-06, "cache_creation_input_token_cost": null, "cache_read_input_token_cost": 1.25e-06, "input_cost_per_character": null, "input_cost_per_token_above_128k_tokens": null, "input_cost_per_token_above_200k_tokens": null, "input_cost_per_query": null, "input_cost_per_second": null, "input_cost_per_audio_token": null, "input_cost_per_token_batches": 1.25e-06, "output_cost_per_token_batches": 5e-06, "output_cost_per_token": 1e-05, "output_cost_per_audio_token": null, "output_cost_per_character": null, "output_cost_per_token_above_128k_tokens": null, "output_cost_per_character_above_128k_tokens": null, "output_cost_per_token_above_200k_tokens": null, "output_cost_per_second": null, "output_cost_per_reasoning_token": null, "output_cost_per_image": null, "output_vector_size": null, "litellm_provider": "openai", "mode": "chat", "supports_system_messages": true, "supports_response_schema": true, "supports_vision": true, "supports_function_calling": true, "supports_tool_choice": true, "supports_assistant_prefill": false, "supports_prompt_caching": true, "supports_audio_input": false, "supports_audio_output": false, "supports_pdf_input": false, "supports_embedding_image_input": false, "supports_native_streaming": null, "supports_web_search": true, "supports_reasoning": false, "search_context_cost_per_query": {"search_context_size_low": 0.03, "search_context_size_medium": 0.035, "search_context_size_high": 0.05}, "tpm": null, "rpm": null, "supported_openai_params": ["frequency_penalty", "logit_bias", "logprobs", "top_logprobs", "max_tokens", "max_completion_tokens", "modalities", "prediction", "n", "presence_penalty", "seed", "stop", "stream", "stream_options", "temperature", "top_p", "tools", "tool_choice", "function_call", "functions", "max_retries", "extra_headers", "parallel_tool_calls", "audio", "response_format", "user"]}}, "additional_usage_values": {"completion_tokens_details": null, "prompt_tokens_details": null}}',
"metadata": '{"applied_guardrails": [], "attempted_fallbacks": null, "original_model_group": null, "batch_models": null, "batch_successful_requests": null, "batch_failed_requests": null, "mcp_tool_call_metadata": null, "vector_store_request_metadata": null, "routing_decision": null, "internal_call_origin": null, "guardrail_information": null, "compression_savings": null, "litellm_gateway_injected_cache": null, "router_metadata": null, "usage_object": {"completion_tokens": 20, "prompt_tokens": 10, "total_tokens": 30, "completion_tokens_details": null, "prompt_tokens_details": null}, "model_map_information": {"model_map_key": "gpt-4o", "model_map_value": {"key": "gpt-4o", "max_tokens": 16384, "max_input_tokens": 128000, "max_output_tokens": 16384, "input_cost_per_token": 2.5e-06, "cache_creation_input_token_cost": null, "cache_read_input_token_cost": 1.25e-06, "input_cost_per_character": null, "input_cost_per_token_above_128k_tokens": null, "input_cost_per_token_above_200k_tokens": null, "input_cost_per_query": null, "input_cost_per_second": null, "input_cost_per_audio_token": null, "input_cost_per_token_batches": 1.25e-06, "output_cost_per_token_batches": 5e-06, "output_cost_per_token": 1e-05, "output_cost_per_audio_token": null, "output_cost_per_character": null, "output_cost_per_token_above_128k_tokens": null, "output_cost_per_character_above_128k_tokens": null, "output_cost_per_token_above_200k_tokens": null, "output_cost_per_second": null, "output_cost_per_reasoning_token": null, "output_cost_per_image": null, "output_vector_size": null, "litellm_provider": "openai", "mode": "chat", "supports_system_messages": true, "supports_response_schema": true, "supports_vision": true, "supports_function_calling": true, "supports_tool_choice": true, "supports_assistant_prefill": false, "supports_prompt_caching": true, "supports_audio_input": false, "supports_audio_output": false, "supports_pdf_input": false, "supports_embedding_image_input": false, "supports_native_streaming": null, "supports_web_search": true, "supports_reasoning": false, "search_context_cost_per_query": {"search_context_size_low": 0.03, "search_context_size_medium": 0.035, "search_context_size_high": 0.05}, "tpm": null, "rpm": null, "supported_openai_params": ["frequency_penalty", "logit_bias", "logprobs", "top_logprobs", "max_tokens", "max_completion_tokens", "modalities", "prediction", "n", "presence_penalty", "seed", "stop", "stream", "stream_options", "temperature", "top_p", "tools", "tool_choice", "function_call", "functions", "max_retries", "extra_headers", "parallel_tool_calls", "audio", "response_format", "user"]}}, "additional_usage_values": {"completion_tokens_details": null, "prompt_tokens_details": null}}',
"cache_key": "Cache OFF",
"spend": 0.00022500000000000002,
"total_tokens": 30,
@ -2961,7 +2961,7 @@ class TestSpendLogsPayload:
"model": "claude-4-sonnet-20250514",
"user": "",
"team_id": "",
"metadata": '{"applied_guardrails": [], "attempted_fallbacks": null, "original_model_group": null, "batch_models": null, "batch_successful_requests": null, "batch_failed_requests": null, "mcp_tool_call_metadata": null, "vector_store_request_metadata": null, "routing_decision": null, "internal_call_origin": null, "guardrail_information": null, "compression_savings": null, "litellm_gateway_injected_cache": null, "usage_object": {"completion_tokens": 503, "prompt_tokens": 2095, "total_tokens": 2598, "completion_tokens_details": null, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}, "model_map_information": {"model_map_key": "claude-4-sonnet-20250514", "model_map_value": {"key": "claude-4-sonnet-20250514", "max_tokens": 128000, "max_input_tokens": 200000, "max_output_tokens": 128000, "input_cost_per_token": 3e-06, "cache_creation_input_token_cost": 3.75e-06, "cache_read_input_token_cost": 3e-07, "input_cost_per_character": null, "input_cost_per_token_above_128k_tokens": null, "input_cost_per_token_above_200k_tokens": null, "input_cost_per_query": null, "input_cost_per_second": null, "input_cost_per_audio_token": null, "input_cost_per_token_batches": null, "output_cost_per_token_batches": null, "output_cost_per_token": 1.5e-05, "output_cost_per_audio_token": null, "output_cost_per_character": null, "output_cost_per_token_above_128k_tokens": null, "output_cost_per_character_above_128k_tokens": null, "output_cost_per_token_above_200k_tokens": null, "output_cost_per_second": null, "output_cost_per_image": null, "output_vector_size": null, "litellm_provider": "anthropic", "mode": "chat", "supports_system_messages": null, "supports_response_schema": true, "supports_vision": true, "supports_function_calling": true, "supports_tool_choice": true, "supports_assistant_prefill": true, "supports_prompt_caching": true, "supports_audio_input": false, "supports_audio_output": false, "supports_pdf_input": true, "supports_embedding_image_input": false, "supports_native_streaming": null, "supports_web_search": false, "supports_reasoning": true, "search_context_cost_per_query": null, "tpm": null, "rpm": null, "supported_openai_params": ["stream", "stop", "temperature", "top_p", "max_tokens", "max_completion_tokens", "tools", "tool_choice", "extra_headers", "parallel_tool_calls", "response_format", "user", "reasoning_effort", "thinking"]}}, "additional_usage_values": {"completion_tokens_details": {"accepted_prediction_tokens": null, "audio_tokens": null, "reasoning_tokens": null, "rejected_prediction_tokens": null, "text_tokens": 503, "image_tokens": null}, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0, "text_tokens": null, "image_tokens": null}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}}',
"metadata": '{"applied_guardrails": [], "attempted_fallbacks": null, "original_model_group": null, "batch_models": null, "batch_successful_requests": null, "batch_failed_requests": null, "mcp_tool_call_metadata": null, "vector_store_request_metadata": null, "routing_decision": null, "internal_call_origin": null, "guardrail_information": null, "compression_savings": null, "litellm_gateway_injected_cache": null, "router_metadata": null, "usage_object": {"completion_tokens": 503, "prompt_tokens": 2095, "total_tokens": 2598, "completion_tokens_details": null, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}, "model_map_information": {"model_map_key": "claude-4-sonnet-20250514", "model_map_value": {"key": "claude-4-sonnet-20250514", "max_tokens": 128000, "max_input_tokens": 200000, "max_output_tokens": 128000, "input_cost_per_token": 3e-06, "cache_creation_input_token_cost": 3.75e-06, "cache_read_input_token_cost": 3e-07, "input_cost_per_character": null, "input_cost_per_token_above_128k_tokens": null, "input_cost_per_token_above_200k_tokens": null, "input_cost_per_query": null, "input_cost_per_second": null, "input_cost_per_audio_token": null, "input_cost_per_token_batches": null, "output_cost_per_token_batches": null, "output_cost_per_token": 1.5e-05, "output_cost_per_audio_token": null, "output_cost_per_character": null, "output_cost_per_token_above_128k_tokens": null, "output_cost_per_character_above_128k_tokens": null, "output_cost_per_token_above_200k_tokens": null, "output_cost_per_second": null, "output_cost_per_image": null, "output_vector_size": null, "litellm_provider": "anthropic", "mode": "chat", "supports_system_messages": null, "supports_response_schema": true, "supports_vision": true, "supports_function_calling": true, "supports_tool_choice": true, "supports_assistant_prefill": true, "supports_prompt_caching": true, "supports_audio_input": false, "supports_audio_output": false, "supports_pdf_input": true, "supports_embedding_image_input": false, "supports_native_streaming": null, "supports_web_search": false, "supports_reasoning": true, "search_context_cost_per_query": null, "tpm": null, "rpm": null, "supported_openai_params": ["stream", "stop", "temperature", "top_p", "max_tokens", "max_completion_tokens", "tools", "tool_choice", "extra_headers", "parallel_tool_calls", "response_format", "user", "reasoning_effort", "thinking"]}}, "additional_usage_values": {"completion_tokens_details": {"accepted_prediction_tokens": null, "audio_tokens": null, "reasoning_tokens": null, "rejected_prediction_tokens": null, "text_tokens": 503, "image_tokens": null}, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0, "text_tokens": null, "image_tokens": null}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}}',
"cache_key": "Cache OFF",
"spend": 0.01383,
"total_tokens": 2598,
@ -3055,7 +3055,7 @@ class TestSpendLogsPayload:
"model": "claude-4-sonnet-20250514",
"user": "",
"team_id": "",
"metadata": '{"applied_guardrails": [], "attempted_fallbacks": 0, "original_model_group": "my-anthropic-model-group", "batch_models": null, "batch_successful_requests": null, "batch_failed_requests": null, "mcp_tool_call_metadata": null, "vector_store_request_metadata": null, "routing_decision": null, "internal_call_origin": null, "guardrail_information": null, "compression_savings": null, "litellm_gateway_injected_cache": null, "usage_object": {"completion_tokens": 503, "prompt_tokens": 2095, "total_tokens": 2598, "completion_tokens_details": null, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}, "model_map_information": {"model_map_key": "claude-4-sonnet-20250514", "model_map_value": {"key": "claude-4-sonnet-20250514", "max_tokens": 128000, "max_input_tokens": 200000, "max_output_tokens": 128000, "input_cost_per_token": 3e-06, "cache_creation_input_token_cost": 3.75e-06, "cache_read_input_token_cost": 3e-07, "input_cost_per_character": null, "input_cost_per_token_above_128k_tokens": null, "input_cost_per_token_above_200k_tokens": null, "input_cost_per_query": null, "input_cost_per_second": null, "input_cost_per_audio_token": null, "input_cost_per_token_batches": null, "output_cost_per_token_batches": null, "output_cost_per_token": 1.5e-05, "output_cost_per_audio_token": null, "output_cost_per_character": null, "output_cost_per_token_above_128k_tokens": null, "output_cost_per_character_above_128k_tokens": null, "output_cost_per_token_above_200k_tokens": null, "output_cost_per_second": null, "output_cost_per_image": null, "output_vector_size": null, "litellm_provider": "anthropic", "mode": "chat", "supports_system_messages": null, "supports_response_schema": true, "supports_vision": true, "supports_function_calling": true, "supports_tool_choice": true, "supports_assistant_prefill": true, "supports_prompt_caching": true, "supports_audio_input": false, "supports_audio_output": false, "supports_pdf_input": true, "supports_embedding_image_input": false, "supports_native_streaming": null, "supports_web_search": false, "supports_reasoning": true, "search_context_cost_per_query": null, "tpm": null, "rpm": null, "supported_openai_params": ["stream", "stop", "temperature", "top_p", "max_tokens", "max_completion_tokens", "tools", "tool_choice", "extra_headers", "parallel_tool_calls", "response_format", "user", "reasoning_effort", "thinking"]}}, "additional_usage_values": {"completion_tokens_details": {"accepted_prediction_tokens": null, "audio_tokens": null, "reasoning_tokens": null, "rejected_prediction_tokens": null, "text_tokens": 503, "image_tokens": null}, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0, "text_tokens": null, "image_tokens": null}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}}',
"metadata": '{"applied_guardrails": [], "attempted_fallbacks": 0, "original_model_group": "my-anthropic-model-group", "batch_models": null, "batch_successful_requests": null, "batch_failed_requests": null, "mcp_tool_call_metadata": null, "vector_store_request_metadata": null, "routing_decision": null, "internal_call_origin": null, "guardrail_information": null, "compression_savings": null, "litellm_gateway_injected_cache": null, "router_metadata": null, "usage_object": {"completion_tokens": 503, "prompt_tokens": 2095, "total_tokens": 2598, "completion_tokens_details": null, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}, "model_map_information": {"model_map_key": "claude-4-sonnet-20250514", "model_map_value": {"key": "claude-4-sonnet-20250514", "max_tokens": 128000, "max_input_tokens": 200000, "max_output_tokens": 128000, "input_cost_per_token": 3e-06, "cache_creation_input_token_cost": 3.75e-06, "cache_read_input_token_cost": 3e-07, "input_cost_per_character": null, "input_cost_per_token_above_128k_tokens": null, "input_cost_per_token_above_200k_tokens": null, "input_cost_per_query": null, "input_cost_per_second": null, "input_cost_per_audio_token": null, "input_cost_per_token_batches": null, "output_cost_per_token_batches": null, "output_cost_per_token": 1.5e-05, "output_cost_per_audio_token": null, "output_cost_per_character": null, "output_cost_per_token_above_128k_tokens": null, "output_cost_per_character_above_128k_tokens": null, "output_cost_per_token_above_200k_tokens": null, "output_cost_per_second": null, "output_cost_per_image": null, "output_vector_size": null, "litellm_provider": "anthropic", "mode": "chat", "supports_system_messages": null, "supports_response_schema": true, "supports_vision": true, "supports_function_calling": true, "supports_tool_choice": true, "supports_assistant_prefill": true, "supports_prompt_caching": true, "supports_audio_input": false, "supports_audio_output": false, "supports_pdf_input": true, "supports_embedding_image_input": false, "supports_native_streaming": null, "supports_web_search": false, "supports_reasoning": true, "search_context_cost_per_query": null, "tpm": null, "rpm": null, "supported_openai_params": ["stream", "stop", "temperature", "top_p", "max_tokens", "max_completion_tokens", "tools", "tool_choice", "extra_headers", "parallel_tool_calls", "response_format", "user", "reasoning_effort", "thinking"]}}, "additional_usage_values": {"completion_tokens_details": {"accepted_prediction_tokens": null, "audio_tokens": null, "reasoning_tokens": null, "rejected_prediction_tokens": null, "text_tokens": 503, "image_tokens": null}, "prompt_tokens_details": {"audio_tokens": null, "cached_tokens": 0, "text_tokens": null, "image_tokens": null}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0}}',
"cache_key": "Cache OFF",
"spend": 0.01383,
"total_tokens": 2598,

69
tests/test_litellm/proxy/spend_tracking/test_spend_tracking_utils.py
View file

@ -2,6 +2,7 @@ import asyncio
import datetime
import json
from datetime import timezone
from collections.abc import Mapping
from typing import Any, Final, cast
from unittest.mock import AsyncMock, MagicMock, patch
@ -3956,3 +3957,71 @@ def test_passthrough_caching_carries_no_injection_marker():
)
metadata = json.loads(payload["metadata"])
assert metadata["litellm_gateway_injected_cache"] is None
def _routed_call_kwargs(model_info: Mapping[str, object]) -> dict[str, object]:
return {
"model": "claude-haiku-4-5",
"custom_llm_provider": "azure_ai",
"litellm_call_id": "router-corr-123",
"litellm_params": {
"metadata": {
"user_api_key": "test-key",
"model_group": "internal-router/gpt-5.4",
"deployment": "azure_ai/claude-haiku-4-5",
"model_info": model_info,
}
},
}
def test_router_metadata_stamped_for_internal_router_model_deployment():
"""A deployment flagged model_info.internal_router_model gets a router_metadata
block correlating the requested model group with the selected deployment."""
payload = get_logging_payload(
kwargs=_routed_call_kwargs({"id": "mi-1", "internal_router_model": True}),
response_obj=litellm.ModelResponse(id="chatcmpl-router-meta", choices=[], usage=litellm.Usage()),
start_time=datetime.datetime.now(timezone.utc),
end_time=datetime.datetime.now(timezone.utc),
)
metadata = json.loads(payload["metadata"])
assert metadata["router_metadata"] == {
"requested_model": "internal-router/gpt-5.4",
"selected_model": "azure_ai/claude-haiku-4-5",
"selected_provider": "azure_ai",
"router_correlation_id": "router-corr-123",
}
def test_router_metadata_absent_without_internal_router_model_flag():
payload = get_logging_payload(
kwargs=_routed_call_kwargs({"id": "mi-1"}),
response_obj=litellm.ModelResponse(id="chatcmpl-unflagged", choices=[], usage=litellm.Usage()),
start_time=datetime.datetime.now(timezone.utc),
end_time=datetime.datetime.now(timezone.utc),
)
metadata = json.loads(payload["metadata"])
assert metadata["router_metadata"] is None
@pytest.mark.parametrize("bucket", ["metadata", "litellm_metadata"])
def test_caller_forged_router_metadata_is_discarded(bucket):
"""The raw request bucket is client-writable and _get_spend_logs_metadata projects
every SpendLogsMetadata key from it, so the server-derived value must overwrite
unconditionally or a caller could plant router provenance the router never produced."""
payload = get_logging_payload(
kwargs={
"model": "gpt-4o-mini",
"litellm_params": {
bucket: {
"user_api_key": "test-key",
"router_metadata": {"requested_model": "forged", "router_correlation_id": "forged-id"},
}
},
},
response_obj=litellm.ModelResponse(id="chatcmpl-forged-router-meta", choices=[], usage=litellm.Usage()),
start_time=datetime.datetime.now(timezone.utc),
end_time=datetime.datetime.now(timezone.utc),
)
metadata = json.loads(payload["metadata"])
assert metadata["router_metadata"] is None

14
tests/test_litellm/proxy/test_common_request_processing.py
View file

@ -1659,29 +1659,29 @@ class TestCommonRequestProcessingHelpers:
async def test_serialize_http_exception_detail_helper(self):
"""Direct unit coverage for the L1 helper across all branches."""
from litellm.proxy.common_request_processing import (
_serialize_http_exception_detail,
serialize_http_exception_detail,
)
import json as _json
assert _serialize_http_exception_detail("plain") == ("plain", None)
assert serialize_http_exception_detail("plain") == ("plain", None)
msg, fields = _serialize_http_exception_detail({"error": "Violated", "extra": "x"})
msg, fields = serialize_http_exception_detail({"error": "Violated", "extra": "x"})
assert msg == "Violated"
assert fields == {"error": "Violated", "extra": "x"}
msg, fields = _serialize_http_exception_detail({"error": {"message": "blocked", "code": "x"}})
msg, fields = serialize_http_exception_detail({"error": {"message": "blocked", "code": "x"}})
assert msg == "blocked"
assert fields == {"error": {"message": "blocked", "code": "x"}}
msg, fields = _serialize_http_exception_detail({"message": "top-level"})
msg, fields = serialize_http_exception_detail({"message": "top-level"})
assert msg == "top-level"
assert fields == {"message": "top-level"}
msg, fields = _serialize_http_exception_detail({"weird": ["a", "b"]})
msg, fields = serialize_http_exception_detail({"weird": ["a", "b"]})
assert msg == _json.dumps({"weird": ["a", "b"]})
assert fields == {"weird": ["a", "b"]}
assert _serialize_http_exception_detail(42) == ("42", None)
assert serialize_http_exception_detail(42) == ("42", None)
async def test_proxy_exception_from_http_exception_helper(self):
"""The shared HTTPException -> ProxyException conversion keeps a clean

29
tests/test_litellm/proxy/test_proxy_logging_hook_detection.py
View file

@ -346,14 +346,14 @@ async def test_post_call_stream_guardrail_keeps_own_iterator_on_chat_completions
@pytest.mark.asyncio
async def test_unified_guardrail_iterator_accepts_explicit_guardrail(monkeypatch):
async def test_unified_guardrail_iterator_accepts_explicit_guardrail():
"""
The dispatch passes each guardrail explicitly instead of through a shared
request_data key, so chaining two unified-routed guardrails cannot drop
all but the last one.
all but the last one. The block fires after the deltas were already
flushed to the client, so it surfaces as a trailing in-stream error frame
rather than a raised HTTPException.
"""
from fastapi import HTTPException
from litellm.proxy.utils import unified_guardrail
guardrail = _content_filter_guardrail("BLOCK")
@ -367,14 +367,19 @@ async def test_unified_guardrail_iterator_accepts_explicit_guardrail(monkeypatch
for chunk in _anthropic_stream_chunks(["the", " zebra runs"]):
yield chunk
with pytest.raises(HTTPException):
async for _ in unified_guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=UserAPIKeyAuth(api_key="sk-1234", request_route="/v1/messages"),
response=fake_stream(),
request_data=request_data,
guardrail_to_apply=guardrail,
):
pass
delivered = []
async for item in unified_guardrail.async_post_call_streaming_iterator_hook(
user_api_key_dict=UserAPIKeyAuth(api_key="sk-1234", request_route="/v1/messages"),
response=fake_stream(),
request_data=request_data,
guardrail_to_apply=guardrail,
):
delivered.append(item)
raw = b"".join(c for c in delivered if isinstance(c, bytes)).decode()
assert "event: error" in raw
assert "guardrail_error" in raw
assert raw.index("guardrail_error") > raw.index(" zebra runs")
@pytest.mark.asyncio

Some files were not shown because too many files have changed in this diff Show more