scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-08 22:21:35 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 78

Merge branch 'litellm_internal_staging' of https://github.com/BerriAI/litellm into litellm_decrease_anys_opus5_r2

Browse source 
# Conflicts:
#	basedpyright-code-budget.json
This commit is contained in:
mateo-berri 2026-08-31 15:05:34 -07:00
commit 4d19a889fb
76 changed files with 3278 additions and 747 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

20
basedpyright-code-budget.json
View file

@ -1,9 +1,9 @@
{
"reportAny": {
"limit": 15468
"limit": 14765
},
"reportArgumentType": {
"limit": 2218
"limit": 2216
},
"reportAssignmentType": {
"limit": 319
@ -24,7 +24,7 @@
"limit": 19
},
"reportExplicitAny": {
"limit": 4846
"limit": 4493
},
"reportFunctionMemberAccess": {
"limit": 7
@ -42,7 +42,7 @@
"limit": 12
},
"reportIndexIssue": {
"limit": 30
"limit": 25
},
"reportInvalidTypeForm": {
"limit": 34
@ -54,10 +54,10 @@
"limit": 0
},
"reportMissingParameterType": {
"limit": 5609
"limit": 5607
},
"reportMissingTypeArgument": {
"limit": 15330
"limit": 15310
},
"reportMissingTypeStubs": {
"limit": 40
@ -105,13 +105,13 @@
"limit": 109
},
"reportUnknownMemberType": {
"limit": 38418
"limit": 38368
},
"reportUnknownParameterType": {
"limit": 19649
"limit": 19633
},
"reportUnknownVariableType": {
"limit": 29987
"limit": 29908
},
"reportUnnecessaryCast": {
"limit": 111
@ -141,6 +141,6 @@
"limit": 543
},
"reportUnusedVariable": {
"limit": 138
"limit": 137
}
}

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

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"

28
litellm/litellm_core_utils/token_counter.py
View file

@ -698,6 +698,26 @@ 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, object],
count_function: TokenCounterFunction,
@ -783,6 +803,12 @@ 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,
@ -812,7 +838,7 @@ 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:

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)

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:

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

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(

267
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",

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

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 ##

18
litellm/proxy/management_endpoints/internal_user_endpoints.py
View file

@ -1027,6 +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
@ -2726,6 +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
@ -2839,6 +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

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:
```

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)],

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

@ -173,7 +173,14 @@ 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 = (
@ -394,6 +417,20 @@ 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,
@ -452,9 +489,13 @@ 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,
),
)
@ -499,15 +540,6 @@ 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:

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

267
model_prices_and_context_window.json
View file

File diff suppressed because it is too large Load diff

16
ruff-strict-budget.json
View file

@ -1,6 +1,6 @@
{
"ANN001": {
"limit": 2993
"limit": 2991
},
"ANN002": {
"limit": 71
@ -12,10 +12,10 @@
"limit": 2002
},
"ANN202": {
"limit": 843
"limit": 841
},
"ANN204": {
"limit": 696
"limit": 694
},
"ANN205": {
"limit": 112
@ -24,7 +24,7 @@
"limit": 133
},
"ANN401": {
"limit": 487
"limit": 387
},
"ASYNC230": {
"limit": 11
@ -117,7 +117,7 @@
"limit": 1
},
"PERF102": {
"limit": 22
"limit": 21
},
"PERF401": {
"limit": 12
@ -168,7 +168,7 @@
"limit": 3
},
"RET504": {
"limit": 174
"limit": 173
},
"RUF012": {
"limit": 239
@ -198,7 +198,7 @@
"limit": 56
},
"SIM102": {
"limit": 312
"limit": 310
},
"SIM103": {
"limit": 119
@ -231,7 +231,7 @@
"limit": 5
},
"TID251": {
"limit": 1096
"limit": 1084
},
"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()

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

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.

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

29
tests/test_litellm/proxy/utils/proxy_logging/test_alerting.py
View file

@ -115,6 +115,35 @@ async def test_budget_alerts_slack_when_slack_alerting(proxy_logging):
assert snapshot == {"type": "user_budget", "user_info_is_callinfo": True, "user_id": "u1"}
@pytest.mark.asyncio
async def test_budget_alerts_webhook_only_forwards_to_slack_alerting_instance(proxy_logging):
proxy_logging.alerting = ["webhook"]
captured: Dict[str, Any] = {}
async def fake_alert(**kwargs):
captured.update(kwargs)
proxy_logging.slack_alerting_instance = MagicMock(budget_alerts=fake_alert)
proxy_logging.email_logging_instance = None
await proxy_logging.budget_alerts(type="user_budget", user_info=_user_info())
snapshot = {
"type": captured["type"],
"user_info_is_callinfo": isinstance(captured["user_info"], CallInfo),
"user_id": captured["user_info"].user_id,
}
assert snapshot == {"type": "user_budget", "user_info_is_callinfo": True, "user_id": "u1"}
@pytest.mark.asyncio
async def test_budget_alerts_email_only_skips_slack_alerting_instance(proxy_logging):
proxy_logging.alerting = ["email"]
proxy_logging.slack_alerting_instance = MagicMock(budget_alerts=AsyncMock())
proxy_logging.email_logging_instance = MagicMock(budget_alerts=AsyncMock())
await proxy_logging.budget_alerts(type="user_budget", user_info=_user_info())
proxy_logging.slack_alerting_instance.budget_alerts.assert_not_called()
proxy_logging.email_logging_instance.budget_alerts.assert_called_once()
@pytest.mark.asyncio
async def test_budget_alerts_soft_budget_with_alert_emails_bypasses_global(proxy_logging):
proxy_logging.alerting = None

19
tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
View file

@ -124,6 +124,25 @@ class TestLiteLLMCompletionResponsesConfig:
assert "extra_field" not in result["file"]
assert "another_field" not in result["file"]
def test_transform_input_file_item_to_file_item_keeps_filename(self):
"""OpenAI rejects file_data with no filename beside it, so dropping it 400s the request"""
result = (
LiteLLMCompletionResponsesConfig._transform_input_file_item_to_file_item(
{
"type": "input_file",
"filename": "report.pdf",
"file_data": "data:application/pdf;base64,JVBERi0=",
}
)
)
assert result == {
"type": "file",
"file": {
"file_data": "data:application/pdf;base64,JVBERi0=",
"filename": "report.pdf",
},
}
def test_transform_input_file_item_to_file_item_with_file_url(self):
"""file_url should be mapped to file_id for downstream URL handling"""
result = (

8
type-discipline-budget.json
View file

@ -1,9 +1,9 @@
{
"LIT001": {
"limit": 22462
"limit": 22403
},
"LIT002": {
"limit": 26800
"limit": 26780
},
"LIT003": {
"limit": 269
@ -27,10 +27,10 @@
"limit": 0
},
"LIT010": {
"limit": 16529
"limit": 16512
},
"LIT011": {
"limit": 5556
"limit": 5537
},
"LIT012": {
"limit": 4495

5
ui/litellm-dashboard/src/components/settings.tsx
View file

@ -522,7 +522,8 @@ const Settings: React.FC<SettingsPageProps> = ({ accessToken, userRole, userID,
<TabsContent value="alerting-types" keepMounted>
<Card className="p-6">
<p className="my-2">
Alerts are only supported for Slack Webhook URLs. Get your webhook urls from{" "}
Alerts are sent to any Slack-compatible incoming webhook URL (Slack, Rocket.Chat, Mattermost, etc.). Get
Slack webhook urls from{" "}
<a href="https://api.slack.com/messaging/webhooks" target="_blank" style={{ color: "blue" }}>
here
</a>
@ -532,7 +533,7 @@ const Settings: React.FC<SettingsPageProps> = ({ accessToken, userRole, userID,
<TableRow>
<TableHead></TableHead>
<TableHead></TableHead>
<TableHead>Slack Webhook URL</TableHead>
<TableHead>Webhook URL (Slack-compatible)</TableHead>
</TableRow>
</TableHeader>

179
ui/litellm-dashboard/src/lib/http/schema.d.ts generated vendored
View file

@ -9700,6 +9700,37 @@ export interface paths {
patch?: never;
trace?: never;
};
"/openai/v1/responses/input_tokens": {
parameters: {
query?: never;
header?: never;
path?: never;
cookie?: never;
};
get?: never;
put?: never;
/**
* Responses Input Tokens
* @description 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>}`
*/
post: operations["responses_input_tokens_openai_v1_responses_input_tokens_post"];
delete?: never;
options?: never;
head?: never;
patch?: never;
trace?: never;
};
"/openai/v1/responses/{response_id}": {
parameters: {
query?: never;
@ -12619,6 +12650,37 @@ export interface paths {
patch?: never;
trace?: never;
};
"/responses/input_tokens": {
parameters: {
query?: never;
header?: never;
path?: never;
cookie?: never;
};
get?: never;
put?: never;
/**
* Responses Input Tokens
* @description 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>}`
*/
post: operations["responses_input_tokens_responses_input_tokens_post"];
delete?: never;
options?: never;
head?: never;
patch?: never;
trace?: never;
};
"/responses/{response_id}": {
parameters: {
query?: never;
@ -15505,7 +15567,7 @@ export interface paths {
* 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
@ -15521,6 +15583,8 @@ export interface paths {
* - 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:
* ```
@ -16210,6 +16274,11 @@ export interface paths {
*
* 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
@ -16241,6 +16310,11 @@ export interface paths {
* Get User Daily Activity Aggregated
* @description 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).
*/
get: operations["get_user_daily_activity_aggregated_user_daily_activity_aggregated_get"];
put?: never;
@ -19182,6 +19256,37 @@ export interface paths {
patch?: never;
trace?: never;
};
"/v1/responses/input_tokens": {
parameters: {
query?: never;
header?: never;
path?: never;
cookie?: never;
};
get?: never;
put?: never;
/**
* Responses Input Tokens
* @description 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>}`
*/
post: operations["responses_input_tokens_v1_responses_input_tokens_post"];
delete?: never;
options?: never;
head?: never;
patch?: never;
trace?: never;
};
"/v1/responses/{response_id}": {
parameters: {
query?: never;
@ -21004,6 +21109,14 @@ export interface paths {
* 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
@ -25182,7 +25295,7 @@ export interface components {
alert_types?: components["schemas"]["AlertType"][] | null;
/**
* Alerting
* @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
*/
alerting?: unknown[] | null;
/**
@ -38418,6 +38531,8 @@ export interface components {
input_cost_per_character?: number | null;
/** Input Cost Per Token */
input_cost_per_token?: number | null;
/** Internal Router Model */
internal_router_model?: boolean | null;
/** Output Cost Per Character */
output_cost_per_character?: number | null;
/** Output Cost Per Token */
@ -51474,6 +51589,26 @@ export interface operations {
};
};
};
responses_input_tokens_openai_v1_responses_input_tokens_post: {
parameters: {
query?: never;
header?: never;
path?: never;
cookie?: never;
};
requestBody?: never;
responses: {
/** @description Successful Response */
200: {
headers: {
[name: string]: unknown;
};
content: {
"application/json": unknown;
};
};
};
};
get_response_openai_v1_responses__response_id__get: {
parameters: {
query?: never;
@ -54438,6 +54573,26 @@ export interface operations {
};
};
};
responses_input_tokens_responses_input_tokens_post: {
parameters: {
query?: never;
header?: never;
path?: never;
cookie?: never;
};
requestBody?: never;
responses: {
/** @description Successful Response */
200: {
headers: {
[name: string]: unknown;
};
content: {
"application/json": unknown;
};
};
};
};
get_response_responses__response_id__get: {
parameters: {
query?: never;
@ -62860,6 +63015,26 @@ export interface operations {
};
};
};
responses_input_tokens_v1_responses_input_tokens_post: {
parameters: {
query?: never;
header?: never;
path?: never;
cookie?: never;
};
requestBody?: never;
responses: {
/** @description Successful Response */
200: {
headers: {
[name: string]: unknown;
};
content: {
"application/json": unknown;
};
};
};
};
get_response_v1_responses__response_id__get: {
parameters: {
query?: never;