fix: normalize response images missing index + guard audio duration overflow

[SabaPivot]

Summary

• Response images missing index field ([Bug]: OpenRouter + google/gemini-3-pro-image-preview: ValidationError - missing index field in images array #22640): Providers like OpenRouter with Gemini image models return images without the index field required by ImageURLListItem. This causes ValidationError when constructing Message. Added _normalize_images() helper to backfill index from enumeration position before passing to Message.

• Audio duration overflow ([Bug]: Audio transcription can be massively overbilled when local soundfile fallback returns invalid duration #22622): libsndfile can report 2^63-1 frames as a sentinel for unknown frame counts in malformed audio files. calculate_request_duration() computes len(audio) / audio.samplerate without sanity checking, producing absurdly large durations (e.g., 192 trillion seconds) used for cost calculation. Added guards for sentinel frame counts and implausible durations (>24 hours), returning None to trigger the existing fallback path.

Test plan

• All 11 llm_response_utils tests pass
• _normalize_images correctly adds index to images missing it
• _normalize_images(None) returns None (no-op for non-image responses)
• Verify OpenRouter Gemini image responses no longer raise ValidationError
• Verify malformed audio files return None duration instead of overflow values

🤖 Generated with claude-flow

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 6, 2026 3:26am

[greptile-apps]

Greptile Summary

This PR fixes two independent bugs: (1) providers like OpenRouter with Gemini image models returning images without the index field required by ImageURLListItem, causing ValidationError on Message construction; and (2) libsndfile reporting 2^63-1 as a sentinel for unknown frame counts in malformed audio files, causing calculate_request_duration() to produce absurdly large cost-calculation inputs.

Key changes:

• _normalize_images() helper in convert_dict_to_response.py backfills a sequential index from enumeration position for any image dict missing the field, before passing the list to Message. The function includes proper type annotations (Optional[List[Dict[str, object]]] -> Optional[List[Dict[str, object]]]).
• calculate_request_duration() in audio_utils/utils.py now guards against: invalid/sentinel frame counts (<= 0 or >= 2^63-1), zero/negative samplerate, and implausible durations exceeding 24 hours — all returning None to trigger the existing fallback cost path.
• No new tests were added for either fix. The two integration-level verification items in the PR checklist remain unchecked ("Verify OpenRouter Gemini image responses no longer raise ValidationError" and "Verify malformed audio files return None duration").

Minor issues:

• A PEP 8 E302 blank-line violation exists before the new _normalize_images function definition (only one blank line after import; two are required).

Confidence Score: 4/5

• The code changes are logically correct and properly guarded. Both fixes are well-scoped with appropriate null/error handling.
• Both code changes are correct and targeted. The type annotations on _normalize_images are present, the samplerate zero-check is in place, and guard conditions properly handle sentinel values and edge cases. Score is 4 (not 5) because: (1) no new tests were added for either bug fix — the integration-level checklist items remain unverified per rule 47800116; (2) a minor PEP 8 E302 linting violation will fail CI; (3) without test evidence, merging relies on code review confidence alone.
• litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py — requires PEP 8 blank-line fix before function definition to pass linting CI.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[convert_to_model_response_object] --> B{choice message\nhas 'images'?}
    B -- No / None --> C[images = None]
    B -- Yes --> D[_normalize_images list]
    D --> E{For each img\nin list}
    E --> F{isinstance dict\nAND 'index' missing?}
    F -- Yes --> G["img = {**img, 'index': i}"]
    F -- No --> H[keep img as-is]
    G --> I[append to normalized]
    H --> I
    I --> E
    E -- done --> J[return normalized list]
    C --> K[Message constructor]
    J --> K

    L[calculate_request_duration] --> M{frames <= 0\nOR frames >= 2^63-1?}
    M -- Yes --> N[return None]
    M -- No --> O{samplerate <= 0?}
    O -- Yes --> N
    O -- No --> P[duration = frames / samplerate]
    P --> Q{duration > 86400s?}
    Q -- Yes --> N
    Q -- No --> R[return duration]

Loading

_{Last reviewed commit: 57e7e70}

[greptile-apps]

Missing type annotations on _normalize_images

The new helper function lacks type hints, which are used consistently throughout the rest of this file (e.g., _safe_convert_created_field, convert_tool_call_to_json_mode). Adding them makes the expected contract explicit and helps static analysis tools catch misuse early.

Suggested change

	def _normalize_images(images):
	"""Normalize image items to include required 'index' field if missing."""
	if images is None:
	return None
	normalized = []
	for i, img in enumerate(images):
	if isinstance(img, dict) and "index" not in img:
	img = {**img, "index": i}
	normalized.append(img)
	return normalized
	def _normalize_images(
	images: Optional[List[ImageURLListItem]],
	) -> Optional[List[ImageURLListItem]]:
	"""Normalize image items to include required 'index' field if missing."""
	if images is None:
	return None
	normalized: List[ImageURLListItem] = []
	for i, img in enumerate(images):
	if isinstance(img, dict) and "index" not in img:
	img = {**img, "index": i}
	normalized.append(img)
	return normalized

_{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!}

[SabaPivot]

Fixed in 57e7e70 — added type annotations to the function signature and local variable.

[greptile-apps]

samplerate zero-check missing

After guarding against invalid frames values, the code still divides by audio.samplerate without verifying it is non-zero. A malformed audio file could report samplerate == 0, which would raise a ZeroDivisionError. In practice the outer except Exception block catches this and returns None, but an explicit guard is safer and communicates intent clearly.

Suggested change

	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
	duration = frames / audio.samplerate
	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

[SabaPivot]

Good catch — fixed in 57e7e70. Added if audio.samplerate <= 0: return None guard before the division.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Missing blank line before top-level function (PEP 8 E302)

PEP 8 requires two blank lines before a top-level function definition. Only one blank line separates the from .get_headers import get_response_headers import on line 46 from _normalize_images on line 48. This will trigger an E302 linting error.

Suggested change

	def _normalize_images(
	def _normalize_images(

To fix, add one more blank line before the function definition.

_{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!}