Backport of #39966 to lts/1.108.x.
Cherry-picked from b39eadcb82 (main).
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Line port of the usage.seconds billing slice onto lts/1.108.x: normalized_audio_duration_seconds and _get_transcription_usage_duration from #35600 (ecb8ef5ca8), and the provider-seconds-first _transcription_duration_for_cost from #50727 (dfb1aac8b1), with their unit tests. The rest of #35600 stays on main.
Backport of #44614 to lts/1.108.x.
Cherry-picked from ba61ecdf57 (main). The line has no splits_stream_timeout on _update_kwargs_with_deployment, so only the _video_call_pin wrapping is added around deployment selection. The line has no litellm/llms/base_llm/videos/async_task_utils.py, so its _id_safe_name dedupe is applied to the dashscope video transformation that holds the same helper on this line.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Backport of #48490 to lts/1.108.x.
Cherry-picked from 76d001a7c0 (main). The line has no wildcard_routed_models helper and no get_models_by_pattern, so only wildcard_route_deployments is changed.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
completion() loaded the MCP gateway handler for any request with tools,
and the handler pulls in mcp 2.x only code, so tool calls failed on
installs with mcp 1.x (google-adk 1.36.1) or no mcp at all
Cherry-picked from 783cf06f31. Only conflict was import context in
litellm/responses/main.py
(cherry picked from commit 783cf06f31)
MiniMax-M3 reached its 2026-11-06 deprecation date in the cost map, which names MiniMax-M3.1-Flash as its successor
(cherry picked from commit 90f7ab6fe6)
The router-model pass-through dropped a Bedrock deployment's guardrailConfig, so converse, converse-stream, invoke and invoke-with-response-stream reached Bedrock unguarded while /v1/chat/completions on the same deployment was blocked. The signer now sends it the way Bedrock reads it, in the body for Converse and as the X-Amzn-Bedrock-Guardrail* headers for Invoke, replacing any guardrail the caller sent
Adapted to the line, which does not have #47040. The guardrail headers are merged through merge_bedrock_invoke_headers without the request metadata header, and the unit tests drop the request metadata assertions and the identity argument that only #47040 needs
Backports test-only changes from main so the line's integration signal stops reporting stale helpers as failures.
Gateway.database_url and Gateway.rows come from #47213 (fc27dfdc7d), and the name keyword on Scenario.model comes from #48917 (ebf84f8649). Only those hunks of tests/integration/_support/client.py are taken. The rest of #47213 gives each owned proxy its own database and Redis, which none of the line's tests need, and the workers field from a later main change is left out.
The MCP progress and cancellation test wraps its upstream in _acknowledge_notifications_after_dispatch, from #49967 (b5952f243a). The helper is already on the line.
The resent-failure retry predicate returned true for BaseException
subclasses such as asyncio.CancelledError, so a cancelled
acompletion_with_retries or aresponses_with_retries call could start
another provider call. Retry only Exception subclasses again, as
tenacity's default did, while still skipping a resent failure.
A transient 429 or 500 on a /v1/messages turn that carries a replayed encrypted
block was retried by an SDK call's num_retries loop, which sent the block again
until every retry was rejected. On async calls the loop then raised the first
transient error, so the turn failed instead of being sent again without the
block. The retry loop now stops at that rejection and passes it on, so the turn
is sent again once without encrypted blocks.
A rejected first try on a sync call also cleared the process-wide
litellm.num_retries, so later calls in the process were no longer retried. That
setting now stays as configured.
A /v1/messages turn whose replayed encrypted block the backend rejects is sent
once more without it. The rejected first attempt no longer fires failure
callbacks or releases its parallel slot and budget reservation, so only the
turn sent again is logged and counted.
Success logs for a Messages client now keep every reasoning item of a
non-streamed turn, matching the reply the client got.
This line's transform_request rewrites the logging call_type to responses, so the non-streamed reply read the Messages-client check after the rewrite and kept only the last reasoning item. The bridge handler now reads it before the request transform and passes it to transform_response, as the streamed path already does.
Backport of #49076 to lts/1.108.x.
Ported from cadfe1d179 (cursor/chatgpt-encrypted-reasoning-messages-1723) onto the line's pass_through/adapters layout.
Backport of #49183 to lts/1.108.x.
Cherry-picked from 1f71ed377f (fix-ollama-chat-stream-tool-call-index).
Adapted to the line. The repair warning tests live in test_litellm_core_utils_prompt_templates_common_utils.py here.
The openai/ Responses path on /v1/messages had no retry, so a chatgpt/ block left in history answered 400 on every later openai/ turn. That path now sends the turn once more without encrypted reasoning, the same way the chat bridge does. The retry matches only invalid_encrypted_content or a "could not be verified" message naming the content, so the exception mapper's own hint no longer triggers it. The xAI web search keep path is backed out, since xAI sends no encrypted reasoning over the bridge.
Backport of 15e6017012 from #49076 to lts/1.108.x, ported by hand onto the pass_through modules.
Co-authored-by: DDT <141296842+Hugo-DDT@users.noreply.github.com>
Backport of #49183 to lts/1.108.x.
Cherry-picked from 5ef6d62cca (fix-ollama-chat-stream-tool-call-index).
Adapted to the line. The repair warning lives in common_utils.py here.
A summary delta carries no item index, so the next item's summary joined the block that already got its signature. A thinking chunk after an indexed signature now opens a new block.
The backported Responses bridge tags each reasoning item's thinking block with its output index, but this line's Anthropic stream wrapper only opened a new block when the block type changed, so a second reasoning item of the same type was dropped. This ports the index check from #46863, which main already has.
