fix: normalize response images missing index + guard audio duration o…#22955
fix: normalize response images missing index + guard audio duration o…#22955krrishdholakia wants to merge 9 commits intomainfrom
Conversation
…verflow (#22950) * feat(spend-logs): add truncation note when error logs are truncated for DB storage (#22936) When the messages or response JSON fields in spend logs are truncated before being written to the database, the truncation marker now includes a note explaining: - This is a DB storage safeguard - Full, untruncated data is still sent to logging callbacks (OTEL, Datadog, etc.) - The MAX_STRING_LENGTH_PROMPT_IN_DB env var can be used to increase the limit Also emits a verbose_proxy_logger.info message when truncation occurs in the request body or response spend log paths. Adds 3 new tests: - test_truncation_includes_db_safeguard_note - test_response_truncation_logs_info_message - test_request_body_truncation_logs_info_message Co-authored-by: Cursor Agent <cursoragent@cursor.com> * Fix admin viewer unable to see all organizations The /organization/list endpoint only checked for PROXY_ADMIN role, causing PROXY_ADMIN_VIEW_ONLY users to fall into the else branch which restricts results to orgs the user is a member of. Use the existing _user_has_admin_view() helper to include both roles. * feat(ui): add Chat UI — ChatGPT-like interface with MCP tools and streaming (#22937) * feat(ui): add chat message and conversation types * feat(ui): add useChatHistory hook for localStorage-backed conversations * feat(ui): add ConversationList sidebar component * feat(ui): add MCPConnectPicker for attaching MCP servers to chat * feat(ui): add ModelSelector dropdown for chat * feat(ui): add ChatInputBar with MCP tool attachment support * feat(ui): add MCPAppsPanel with list/detail view for MCP servers * feat(ui): add ChatMessages component; remove auto-scrollIntoView that caused scroll-lock bypass * feat(ui): add ChatPage — ChatGPT-like UI with scroll lock, MCP tools, streaming * feat(ui): add /chat route wired to ChatPage * feat(ui): remove chat from leftnav — chat accessible via navbar button * feat(ui): add Chat button to top navbar * feat(ui): add dismissible Chat UI announcement banner to Playground page * feat(proxy): add Chat UI link to Swagger description * feat(ui): add react-markdown and syntax-highlighter deps for chat UI * fix(ui): replace missing BorderOutlined import with inline stop icon div * fix(ui): apply remark-gfm plugin to ReactMarkdown for GFM support * fix(ui): remove unused isEvenRow variable in MCPAppsPanel * fix(ui): add ellipsis when truncating conversation title * fix(ui): wire search button to chats view; remove non-functional keyboard hint * fix(ui): use serverRootPath in navbar chat link for sub-path deployments * fix(ui): remove unused ChatInputBar and ModelSelector files * fix(ui): correct grid bottom-border condition for odd server count * fix(chat): move localStorage writes out of setConversations updater (React purity) * fix(chat): fix stale closure in handleEditAndResend - compute history before async state update * fix(chat): fix 4 issues in ChatMessages - array redaction, clipboard error, inline detection, remove unused ref * docs: add PayGo/priority cost tracking for Gemini Vertex AI - Add PayGo / Priority Cost Tracking section to Vertex AI provider docs - Document trafficType to service_tier mapping (ON_DEMAND_PRIORITY, FLEX, etc.) - Add service tier cost keys to custom pricing docs - Add provider-specific cost tracking note to spend tracking overview Made-with: Cursor * fix: normalize response images missing index + guard audio duration overflow 1. convert_dict_to_response.py (#22640): Providers like OpenRouter/Gemini return images without the required `index` field, causing pydantic ValidationError when constructing Message. Added _normalize_images() to backfill index from enumeration position. 2. audio_utils/utils.py (#22622): libsndfile can report 2^63-1 frames for malformed audio files, causing astronomically large duration values used for cost calculation. Added guards for sentinel frame counts and implausible durations (>24h). Co-Authored-By: claude-flow <ruv@ruv.net> * fix: add type annotations to _normalize_images + guard samplerate==0 Address review feedback: - Add type hints to _normalize_images() for consistency with codebase - Guard against samplerate <= 0 to prevent ZeroDivisionError on malformed audio files Co-Authored-By: claude-flow <ruv@ruv.net> --------- Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com> Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Ryan Crabbe <rcrabbe@berkeley.edu> Co-authored-by: ryan-crabbe <128659760+ryan-crabbe@users.noreply.github.com> Co-authored-by: Ishaan Jaff <ishaanjaffer0324@gmail.com> Co-authored-by: Sameer Kankute <sameer@berri.ai> Co-authored-by: claude-flow <ruv@ruv.net>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Giulio Leone seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. You have signed the CLA already but the status is still pending? Let us recheck it. |
Greptile SummaryThis PR bundles several independent bug fixes: audio duration overflow guards, Anthropic empty-text-block sanitization for the Key findings:
Confidence Score: 4/5
Last reviewed commit: a5e10f9 |
…22910) Fix malformed URL construction in OllamaConfig.get_model_info() when api_base already contains an endpoint path (e.g., /api/generate). Before this fix, when api_base was passed with an endpoint path already appended (which happens through the completion flow via get_complete_url()), get_model_info() would naively append /api/show, resulting in URLs like: http://server:11434/api/generate/api/show (404 error) This fix strips known endpoint paths (/api/generate, /api/chat, /api/embed) before appending /api/show, following the same defensive pattern used elsewhere in the codebase (e.g., embedding handler, chat transformation). Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a58d859217
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| {/* Sidebar nav buttons */} | ||
| <div style={{ padding: "0 8px 4px", flexShrink: 0 }}> | ||
| {sidebarNavItem(<EditOutlined />, "New chat", () => router.push("/chat"))} |
There was a problem hiding this comment.
This uses an absolute "/chat" path, which drops the dashboard base prefix used in proxy deployments (/ui, plus optional server_root_path). In production (NEXT_PUBLIC_BASE_URL="ui/"), this can navigate users outside the mounted UI and produce 404s or non-dashboard routes when starting a new chat or recovering from stale IDs; these navigations should be built with the same base-path helper used elsewhere (/ui/chat with server root prefix).
Useful? React with 👍 / 👎.
| // Toggle ON — verify tools are reachable first | ||
| setTogglingOn((prev) => new Set(prev).add(serverName)); | ||
| try { | ||
| const result = await listMCPTools(accessToken, serverName); |
There was a problem hiding this comment.
The tool-validation call passes serverName (derived from server_name/alias) into listMCPTools, but that endpoint is queried as ?server_id=... and access checks are done against allowed server IDs. When display name differs from server_id, toggle-on will incorrectly fail with "could not load tools" and the server cannot be connected for chat. Use server.server_id for the validation call and keep display names only for UI labels.
Useful? React with 👍 / 👎.
* feat(ovhcloud): Add support of responses API * Edit from greptile recommendation * fix support of tool calling
* docs: add AgentField to integrations index * docs: add AgentField tutorial page * docs: add AgentField to Agent SDKs sidebar
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| frames = len(audio) | ||
| # Guard against sentinel/invalid frame counts (e.g., 2^63-1 from libsndfile) | ||
| if frames <= 0 or frames >= 2**63 - 1: | ||
| return None | ||
| if audio.samplerate <= 0: | ||
| return None | ||
| duration = frames / audio.samplerate | ||
| # Reject implausible durations (> 24 hours) | ||
| if duration > 86400: | ||
| return None |
There was a problem hiding this comment.
No unit tests for the sentinel guard and samplerate fix
The two new guards (frames >= 2**63 - 1 and samplerate <= 0) are described as the primary bug fix in the PR title, yet no corresponding test was added in tests/test_litellm/litellm_core_utils/test_audio_utils.py. The project's own pre-submission checklist states "Adding at least 1 test is a hard requirement". Without tests it's difficult to verify the sentinel-value path returns None and that the samplerate guard prevents a ZeroDivisionError.
| def _normalize_images( | ||
| images: Optional[List[Dict[str, object]]], | ||
| ) -> Optional[List[Dict[str, object]]]: | ||
| """Normalize image items to include required 'index' field if missing.""" | ||
| if images is None: | ||
| return None | ||
| normalized: List[Dict[str, object]] = [] | ||
| for i, img in enumerate(images): | ||
| if isinstance(img, dict) and "index" not in img: | ||
| img = {**img, "index": i} | ||
| normalized.append(img) | ||
| return normalized |
There was a problem hiding this comment.
No unit tests for _normalize_images
This is the other primary bug fix in the PR title (OpenRouter/Gemini images missing index causing pydantic ValidationError), yet no test was added for _normalize_images in tests/test_litellm/. A simple test asserting that:
- a list of dicts without
indexgets sequential indices backfilled - a list of dicts already having
indexis returned unchanged Noneis returned as-is
…would greatly increase confidence in the fix and protect against regressions.
| from litellm.constants import ( | ||
| LITELLM_TRUNCATED_PAYLOAD_FIELD, | ||
| LITELLM_TRUNCATION_DB_SAFEGUARD_NOTE, | ||
| ) |
There was a problem hiding this comment.
Redundant local import inside function body
LITELLM_TRUNCATED_PAYLOAD_FIELD and LITELLM_TRUNCATION_DB_SAFEGUARD_NOTE are now imported at module level (lines 14–17) and again inside _sanitize_request_body_for_spend_logs_payload. The inner import is harmless (Python deduplicates module imports), but it's inconsistent and clutters the function. The local import can be removed.
| from litellm.constants import ( | |
| LITELLM_TRUNCATED_PAYLOAD_FIELD, | |
| LITELLM_TRUNCATION_DB_SAFEGUARD_NOTE, | |
| ) | |
| if visited is None: |
Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
…22933) The `/v1/messages` endpoint does not sanitize empty text content blocks from messages, causing 400 errors ("text content blocks must be non-empty") in multi-turn tool-use conversations. Claude's API returns assistant messages with empty text blocks alongside `tool_use` blocks (e.g., `{"type": "text", "text": ""}`). While the API returns these, it rejects them when sent back in subsequent requests. The `/v1/chat/completions` path already handles this via `process_empty_text_blocks()` in `factory.py`, but the `/v1/messages` path was missing equivalent sanitization. Fix: Add `_sanitize_anthropic_messages()` at the entry point of the `/v1/messages` handler, before any routing. This strips empty/whitespace text blocks from message content arrays while preserving non-empty text and other block types (`tool_use`, `tool_result`, etc.). The function does not leave content arrays empty to avoid a different validation error. This affects all providers routed through `/v1/messages` (Anthropic direct, Bedrock Converse, Azure, etc.) and is particularly important for Claude Code / Agent SDK conversations which frequently produce these patterns. Fixes #22930
| if not isinstance(content, list): | ||
| continue | ||
|
|
||
| # Filter out empty text blocks, keeping non-empty text and other types. | ||
| # Use `(... or "")` to guard against None text values. | ||
| filtered = [ | ||
| block | ||
| for block in content | ||
| if not ( | ||
| isinstance(block, dict) | ||
| and block.get("type") == "text" | ||
| and not (block.get("text") or "").strip() | ||
| ) | ||
| ] | ||
|
|
||
| # Only update if we actually removed something. | ||
| # Avoid mutating the caller's dicts — create a shallow copy. | ||
| if len(filtered) < len(content): | ||
| if len(filtered) > 0: | ||
| messages[i] = {**message, "content": filtered} | ||
| else: | ||
| # All blocks were empty text — replace with a continuation | ||
| # message rather than leaving empty blocks that trigger 400 | ||
| # errors. Matches behavior of process_empty_text_blocks() | ||
| # in factory.py. | ||
| messages[i] = { | ||
| **message, | ||
| "content": [{"type": "text", "text": DEFAULT_ASSISTANT_CONTINUE_MESSAGE.get("content", "Please continue.")}], |
There was a problem hiding this comment.
In-place mutation of the input messages list
The function docstring comment says "Avoid mutating the caller's dicts — create a shallow copy", but the function actually mutates the input messages list itself via messages[i] = .... Only the individual message dicts get shallow copies — the outer list is still modified in-place.
While each updated message dict is a new object ({**message, ...}), lines like:
messages[i] = {**message, "content": filtered}…write back into the original list. Any caller that holds a reference to the same list object (e.g. for retry logic, logging, or span attributes) will observe the sanitised content unexpectedly.
A safe alternative is to build a new list rather than modifying in place:
def _sanitize_anthropic_messages(messages: List[Dict]) -> List[Dict]:
result = []
for message in messages:
content = message.get("content")
if not isinstance(content, list):
result.append(message)
continue
filtered = [
block
for block in content
if not (
isinstance(block, dict)
and block.get("type") == "text"
and not (block.get("text") or "").strip()
)
]
if len(filtered) < len(content):
if len(filtered) > 0:
result.append({**message, "content": filtered})
else:
result.append({
**message,
"content": [{"type": "text", "text": DEFAULT_ASSISTANT_CONTINUE_MESSAGE.get("content", "Please continue.")}],
})
else:
result.append(message)
return result| _request_body = _sanitize_request_body_for_spend_logs_payload(_request_body) | ||
| _request_body_json_str = json.dumps(_request_body, default=str) | ||
| if LITELLM_TRUNCATED_PAYLOAD_FIELD in _request_body_json_str: | ||
| verbose_proxy_logger.info( | ||
| "Spend Log: request body was truncated before storing in DB. %s", |
There was a problem hiding this comment.
Substring check can produce false-positive truncation log entries
LITELLM_TRUNCATED_PAYLOAD_FIELD equals the string "litellm_truncated". The check:
if LITELLM_TRUNCATED_PAYLOAD_FIELD in _request_body_json_str:…is a plain substring search across the entire serialised JSON. If any key or value in the request body legitimately contains the text "litellm_truncated" (e.g. a prompt that discusses LiteLLM internals, or a model name / metadata field), the info log will fire even though no actual truncation occurred. The same pattern appears in _get_response_for_spend_logs_payload (line ~887).
A more robust approach is to have _sanitize_request_body_for_spend_logs_payload return a flag or a lightweight wrapper indicating whether truncation happened, and base the log decision on that flag rather than a substring scan of the full JSON string.
| from litellm.types.router import GenericLiteLLMParams | ||
| from litellm.types.utils import LlmProviders | ||
| from litellm.utils import get_model_info | ||
|
|
There was a problem hiding this comment.
Missing blank lines before class definition
PEP 8 requires two blank lines before a top-level class definition. There is currently zero blank lines between the last import and class OVHCloudResponsesAPIConfig. The same issue exists in convert_dict_to_response.py where _normalize_images is separated from the preceding import by only one blank line.
| from litellm.utils import get_model_info | |
| class OVHCloudResponsesAPIConfig(OpenAIResponsesAPIConfig): |
…ent provider prefix (#22906) * fix(cost_calc): update custom_llm_provider when base_model has different provider When base_model carries a provider prefix that differs from the deployment provider (e.g. base_model='gemini/gemini-2.0-flash' on an anthropic/ deployment), the custom_llm_provider was not updated, causing cost_per_token to build an invalid model key and return 0. After _select_model_name_for_cost_calc resolves the model name from base_model, extract the provider prefix and update custom_llm_provider so the downstream cost lookup uses the correct provider. Fixes #22257 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix(cost_calc): gate base_model provider override on custom_pricing and diff check - Skip custom_llm_provider override when custom_pricing=True (base_model unused) - Only override when extracted provider differs from current custom_llm_provider - Add direct completion_cost unit test for cross-provider base_model - Add same-provider no-regression test (e.g. openai/gpt-4o on openai deployment) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: guard hidden_params override when provider already overridden by base_model - Add _provider_overridden flag to prevent hidden_params from undoing base_model fix - Add direct unit test verifying hidden_params doesn't override extracted provider Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: gate reasoning_auto_summary injection on supports_reasoning() When reasoning_auto_summary is globally enabled, the reasoning param was injected unconditionally for all models including non-reasoning ones (e.g. gpt-4o-mini), causing OpenAI API errors. Now gated on supports_reasoning(model, custom_llm_provider) check. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: revert unrelated transformation.py changes causing TypeError The transformation.py changes pass model= and custom_llm_provider= kwargs to _map_optional_params_to_responses_api_request() which only accepts (self, optional_params, responses_api_request) — causing a TypeError at runtime for every Responses API request. Reverted to upstream version; cost_calculator.py fix is unaffected. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: giulio-leone <6887247+giulio-leone@users.noreply.github.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
| setCopied(true); | ||
| setTimeout(() => setCopied(false), 2000); | ||
| }).catch(() => { | ||
| message.error("Failed to copy to clipboard"); |
There was a problem hiding this comment.
message is not imported from antd — this will cause a ReferenceError at runtime when navigator.clipboard.writeText fails (e.g., in non-HTTPS contexts or with restricted clipboard permissions).
Add message to the antd import:
| message.error("Failed to copy to clipboard"); | |
| import { Collapse, Tooltip, message } from "antd"; |
Alternatively, handle the error with console.error() instead.
| if len(filtered) > 0: | ||
| messages[i] = {**message, "content": filtered} | ||
| else: | ||
| # All blocks were empty text — replace with a continuation | ||
| # message rather than leaving empty blocks that trigger 400 | ||
| # errors. Matches behavior of process_empty_text_blocks() | ||
| # in factory.py. | ||
| messages[i] = { | ||
| **message, | ||
| "content": [{"type": "text", "text": DEFAULT_ASSISTANT_CONTINUE_MESSAGE.get("content", "Please continue.")}], | ||
| } | ||
|
|
There was a problem hiding this comment.
The function mutates the input messages list in-place via messages[i] = ..., despite the comment at line 91 suggesting immutability ("Avoid mutating the caller's dicts"). While each dict is shallow-copied, the list structure itself is modified, which violates the documented contract.
To ensure immutability and prevent bugs in callers that hold references to the original list, build a new list instead:
def _sanitize_anthropic_messages(messages: List[Dict]) -> List[Dict]:
"""..."""
result = []
for message in messages:
content = message.get("content")
if not isinstance(content, list):
result.append(message)
continue
filtered = [
block
for block in content
if not (
isinstance(block, dict)
and block.get("type") == "text"
and not (block.get("text") or "").strip()
)
]
if len(filtered) < len(content):
if len(filtered) > 0:
result.append({**message, "content": filtered})
else:
result.append({
**message,
"content": [{"type": "text", "text": DEFAULT_ASSISTANT_CONTINUE_MESSAGE.get("content", "Please continue.")}],
})
else:
result.append(message)
return resultThe call site already reassigns (messages = _sanitize_anthropic_messages(messages)), so this change is safe.
| def _normalize_images( | ||
| images: Optional[List[Dict[str, object]]], | ||
| ) -> Optional[List[Dict[str, object]]]: | ||
| """Normalize image items to include required 'index' field if missing.""" | ||
| if images is None: | ||
| return None | ||
| normalized: List[Dict[str, object]] = [] | ||
| for i, img in enumerate(images): | ||
| if isinstance(img, dict) and "index" not in img: | ||
| img = {**img, "index": i} | ||
| normalized.append(img) | ||
| return normalized |
There was a problem hiding this comment.
_normalize_images is dead code — never called
This function was added by the PR but is never invoked anywhere in the codebase. A grep of the whole repo confirms the only occurrence is its own definition at line 49.
The actual image-index fix is handled by the pre-existing _normalize_images_for_message() (lines 70–85), which was already present before this PR and is already called at line 627:
images=_normalize_images_for_message(
choice["message"].get("images", None)
),The two functions are nearly identical in behavior. Because _normalize_images is never reached by any call path, the stated bug fix is entirely carried by the pre-existing helper — not by this new function. Either remove this function or wire it up to an actual call site.
| from litellm.proxy.management_endpoints.jwt_key_mapping_endpoints import ( | ||
| router as jwt_key_mapping_router, | ||
| ) | ||
| from litellm.proxy.management_endpoints.jwt_key_mapping_endpoints import ( | ||
| router as jwt_key_mapping_router, | ||
| ) |
There was a problem hiding this comment.
Duplicate import of jwt_key_mapping_router
Lines 377–379 and 380–382 import the same name from the same module twice. This is harmless at runtime (Python deduplicates module imports) but is a lint error and adds noise.
Remove the second import block (lines 380–382); the first import at lines 377–379 is sufficient.
| supports_function_calling: Optional[bool] = None | ||
| try: | ||
| model_info = get_model_info(model, custom_llm_provider="ovhcloud") | ||
| supports_function_calling = model_info.get( | ||
| "supports_function_calling", False | ||
| ) | ||
| except Exception as e: | ||
| verbose_logger.debug(f"Error getting supported OpenAI params: {e}") | ||
| pass | ||
|
|
||
| if supports_function_calling is not True: | ||
| verbose_logger.debug( | ||
| "You can see our models supporting function_calling in our catalog: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/ " | ||
| ) | ||
| # Remove tool-related params for models that don't support function calling | ||
| for param in ("tools", "tool_choice"): | ||
| if param in supported_params: | ||
| supported_params.remove(param) | ||
|
|
||
| return supported_params |
There was a problem hiding this comment.
Debug log fires for both unknown and confirmed non-function-calling models
supports_function_calling is initialised to None. When get_model_info succeeds and returns False (model is confirmed not to support function calling), the condition at line 50 (if supports_function_calling is not True) evaluates to True, and the debug message is emitted.
This means the message is logged in two scenarios:
- When
get_model_inforaised an exception (supports_function_callingis stillNone) — warrants the "check the catalog" hint - When the model is confirmed not to support function calling (
valueisFalse) — expected behavior with no actionable guidance needed
Only case 1 should emit the hint. Consider separating the logic:
if supports_function_calling is None:
verbose_logger.debug(
"Could not determine function calling support; "
"see our catalog: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/"
)
if supports_function_calling is not True:
for param in ("tools", "tool_choice"):
if param in supported_params:
supported_params.remove(param)10 participants
…verflow (#22950)
When the messages or response JSON fields in spend logs are truncated before being written to the database, the truncation marker now includes a note explaining:
Also emits a verbose_proxy_logger.info message when truncation occurs in the request body or response spend log paths.
Adds 3 new tests:
The /organization/list endpoint only checked for PROXY_ADMIN role, causing PROXY_ADMIN_VIEW_ONLY users to fall into the else branch which restricts results to orgs the user is a member of. Use the existing _user_has_admin_view() helper to include both roles.
feat(ui): add Chat UI — ChatGPT-like interface with MCP tools and streaming (feat(ui): add Chat UI — ChatGPT-like interface with MCP tools and streaming #22937)
feat(ui): add chat message and conversation types
feat(ui): add useChatHistory hook for localStorage-backed conversations
feat(ui): add ConversationList sidebar component
feat(ui): add MCPConnectPicker for attaching MCP servers to chat
feat(ui): add ModelSelector dropdown for chat
feat(ui): add ChatInputBar with MCP tool attachment support
feat(ui): add MCPAppsPanel with list/detail view for MCP servers
feat(ui): add ChatMessages component; remove auto-scrollIntoView that caused scroll-lock bypass
feat(ui): add ChatPage — ChatGPT-like UI with scroll lock, MCP tools, streaming
feat(ui): add /chat route wired to ChatPage
feat(ui): remove chat from leftnav — chat accessible via navbar button
feat(ui): add Chat button to top navbar
feat(ui): add dismissible Chat UI announcement banner to Playground page
feat(proxy): add Chat UI link to Swagger description
feat(ui): add react-markdown and syntax-highlighter deps for chat UI
fix(ui): replace missing BorderOutlined import with inline stop icon div
fix(ui): apply remark-gfm plugin to ReactMarkdown for GFM support
fix(ui): remove unused isEvenRow variable in MCPAppsPanel
fix(ui): add ellipsis when truncating conversation title
fix(ui): wire search button to chats view; remove non-functional keyboard hint
fix(ui): use serverRootPath in navbar chat link for sub-path deployments
fix(ui): remove unused ChatInputBar and ModelSelector files
fix(ui): correct grid bottom-border condition for odd server count
fix(chat): move localStorage writes out of setConversations updater (React purity)
fix(chat): fix stale closure in handleEditAndResend - compute history before async state update
fix(chat): fix 4 issues in ChatMessages - array redaction, clipboard error, inline detection, remove unused ref
docs: add PayGo/priority cost tracking for Gemini Vertex AI
Made-with: Cursor
convert_dict_to_response.py ([Bug]: OpenRouter + google/gemini-3-pro-image-preview: ValidationError - missing
indexfield in images array #22640): Providers like OpenRouter/Gemini return images without the requiredindexfield, causing pydantic ValidationError when constructing Message. Added _normalize_images() to backfill index from enumeration position.audio_utils/utils.py ([Bug]: Audio transcription can be massively overbilled when local soundfile fallback returns invalid duration #22622): libsndfile can report 2^63-1 frames for malformed audio files, causing astronomically large duration values used for cost calculation. Added guards for sentinel frame counts and implausible durations (>24h).
Address review feedback:
Relevant issues
Pre-Submission checklist
Please complete all items before asking a LiteLLM maintainer to review your PR
tests/test_litellm/directory, Adding at least 1 test is a hard requirement - see detailsmake test-unit@greptileaiand received a Confidence Score of at least 4/5 before requesting a maintainer reviewCI (LiteLLM team)
Branch creation CI run
Link:
CI run for the last commit
Link:
Merge / cherry-pick CI run
Links:
Type
🆕 New Feature
🐛 Bug Fix
🧹 Refactoring
📖 Documentation
🚄 Infrastructure
✅ Test
Changes